Struct GptOssRotaryEmbedding 

pub struct GptOssRotaryEmbedding { /* private fields */ }

GPT-OSS style rotary embedding with YARN scaling support. Uses chunked/GPT-NeoX style rotation and applies attention scaling.

Implementations

impl GptOssRotaryEmbedding

pub fn new( base: f64, head_dim: usize, max_position_embeddings: usize, factor: f64, original_max_position_embeddings: usize, beta_fast: f64, beta_slow: f64, truncate: bool, device: &Device, dtype: DType, ) -> Result<Self>

Create a new GPT-OSS rotary embedding with YARN scaling.

Arguments

• base - Base frequency for RoPE
• head_dim - Dimension of each attention head
• max_position_embeddings - Maximum sequence length
• factor - YARN scaling factor
• original_max_position_embeddings - Original max positions before scaling
• beta_fast - YARN beta_fast parameter
• beta_slow - YARN beta_slow parameter
• truncate - Whether to truncate correction dimensions
• device - Device to create tensors on
• dtype - Data type for the embeddings

pub fn forward( &self, q: &Tensor, k: &Tensor, seqlen_offsets: &[usize], ) -> Result<(Tensor, Tensor)>

Trait Implementations

impl Clone for GptOssRotaryEmbedding

fn clone(&self) -> GptOssRotaryEmbedding

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for GptOssRotaryEmbedding

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

Formats the value using the given formatter.