pub fn apply_llama_rope(
x: Tensor<4>,
cos_table: Tensor<2>,
sin_table: Tensor<2>,
start_pos: usize,
) -> Tensor<4>Expand description
Llama-style rotary embedding apply on [b, heads, seq, head_dim]
with prebuilt [max_seq, head_dim] cos/sin tables (positions start..).
rotate_half(x) = cat([-x[half:], x[:half]]); out = x*cos + rot*sin.
Shared by Chronos, Sundial, Toto (all use this exact form).