Struct Embedding 

pub struct Embedding {
    pub wte_chunks: Vec<Tensor>,
    pub chunk_rows: usize,
    pub wpe: Tensor,
}

Token + positional embedding table. The token table is stored row-chunked so no single GPU binding exceeds the device limit (see roadmap pitfalls); on CPU it is a single chunk.

Fields

wte_chunks: Vec<Tensor>

Row chunks of the [vocab, n_embd] token table.

chunk_rows: usize

Rows per chunk (the last chunk may be smaller).

wpe: Tensor

[n_ctx, n_embd]

Implementations

impl Embedding

pub fn from_host_wte( wte: &[f32], vocab: usize, n_embd: usize, wpe: Tensor, device: &Device, ) -> Result<Self>

Split a host-side [vocab, n_embd] table into device chunks that each fit in a single binding on device.

pub fn forward(&self, ids: &Tensor, pos: usize) -> Result<Tensor>

ids: u32 tensor [t]; pos is the absolute position of ids[0].