Struct RotaryEmbedding 

pub struct RotaryEmbedding { /* private fields */ }

Precomputed cos/sin tables for every position up to max_position, so RotaryEmbedding::apply is a table lookup per (position, pair) rather than a cos/sin call — cheap enough to matter once this runs once per token per layer per head during decode.

Implementations

impl RotaryEmbedding

pub fn new(rope_dim: usize, theta: f32, max_position: usize) -> Self

Precomputes rotation tables for every position in 0..max_position.

rope_dim must be even (each pair needs two dimensions; see crate::config::QwenConfig::from_metadata, which rejects an odd rope_dimension_count before a RotaryEmbedding is ever built).

pub fn apply(&self, x: &Tensor, positions: &[usize]) -> Result<Tensor>

Rotates x, shaped [n_heads, seq, head_dim], in place per position. positions[s] is the absolute sequence position of x’s s-th row — not necessarily s itself, since a KV-cache decode step’s single query token sits at cache_len, not at 0 (see crate::kv_cache::KvCache).

Only the leading rope_dim elements of each head are rotated; head_dim - rope_dim trailing elements (when rope_dim < head_dim) pass through unchanged, matching llama.cpp’s partial-rotary support.

Errors

Error::ShapeMismatch if x is not rank 3, if its last dimension is smaller than rope_dim, or if positions.len() does not match x’s sequence dimension.