Enum RotationSeq 

pub enum RotationSeq {
    Angle {
        cum_bsha: Tensor<4>,
        rope_dim: usize,
        rotate_pairwise: bool,
    },
    Quaternion {
        cum_bshj4: Tensor<5>,
    },
}

Per-token cumulative rotation of one forward call — the sequence-resolved counterpart of the cache’s final RotationState, exported by rotate_bc_forward (it is already materialised there) for the physical-frame state moments: the per-token states must be de-rotated by the inverse of the rotation absorbed into B/C (see Self::derotate_states). Carries its own application metadata so a consumer cannot re-derive a mismatched pairing layout.

Variants

Angle

Abelian cumulative RoPE angles.

Fields

cum_bsha: Tensor<4>

Cumulative angles [batch, sequence, nheads, num_rope_angles] (un-wrapped — every consumer is 2π-periodic).

rope_dim: usize

Number of rotated B/C entries (0 ⇒ the rotation is the identity).

rotate_pairwise: bool

Interleaved/NeoX (SISO) vs half-and-half/GPT-J (MIMO) pairing — exactly the flag apply_rope_partial was called with.

Quaternion

Quaternion cumulative rotation [batch, sequence, nheads, blocks, 4]; rotates the first 4·blocks state-rank entries as consecutive blocks.

Fields

cum_bshj4: Tensor<5>

Cumulative unit quaternions (the quat_cumprod output).

Implementations

impl RotationSeq

pub fn detached(&self) -> Self

A value-identical copy detached from any autodiff graph.

pub fn pad_to(self, padded_len: usize) -> Self

Zero-pad the sequence axis to padded_len (the SSD chunk padding). Pad values are never read un-masked: the moments mask excludes pad positions before accumulation.

pub fn derotate_states( &self, states_blhpr: Tensor<5>, start: usize, ) -> Tensor<5>

Rotate the state_rank axis of per-token cache-frame states back to the physical frame — the inverse of the rotation rotate_bc_forward absorbed into B/C, so the result cₜ is what raw, un-rotated C reads (yₜ = C̃ₜᵀh̃ₜ = Cₜᵀcₜ):

• Angle: B/C were rotated by +θₜ, so the state gets R(−θₜ) (negated angles).
• Quaternion: B/C were left-multiplied by conj(Qₜ), so the state gets Qₜ un-conjugated.

start is the offset of this chunk of states within the rotation’s sequence axis.

Shapes

• states_blhpr: [batch, len, nheads, per_head_dim, state_rank]
• output: same shape.

Trait Implementations

impl Clone for RotationSeq

fn clone(&self) -> RotationSeq

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for RotationSeq

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.