Struct LayaGraph 

pub struct LayaGraph {
    pub tok_embeddings: W,
    pub embed_norm: W,
    pub layers: Vec<EncoderLayer>,
    pub final_norm: W,
    pub type_emb: W,
    pub head: Vec<HeadLayer>,
    pub scorer_norm: Affine,
    pub scorer_in: Affine,
    pub scorer_out: Affine,
    pub act_in: Affine,
    pub act_out: Affine,
    pub temperature: W,
}

The whole compat graph with every weight bound.

Fields

tok_embeddings: W

[vocab, d].

embed_norm: W

[d].

layers: Vec<EncoderLayer>

Encoder layers in order.

final_norm: W

[d].

type_emb: W

[3, d], one row per question type.

head: Vec<HeadLayer>

Decision head layers in order.

scorer_norm: Affine

LayerNorm, scorer.0.

scorer_in: Affine

scorer.1, [d, d].

scorer_out: Affine

scorer.3, [1, d].

act_in: Affine

act_head.0, [256, d + 4].

act_out: Affine

act_head.2, [2, 256].

temperature: W

[3], the calibrated temperatures stored in the checkpoint.

Implementations

impl LayaGraph

pub fn bind(spec: &LayaSpec, tensors: &Tensors) -> Result<Self, Error>

Binds every weight of the graph to tensors, checking names, shapes and dtypes.

Errors

Error::Mismatch listing every missing tensor, extra tensor, wrong shape and non float dtype, in that order.

Panics

Never. Binding looks up only names the check above has found.

pub fn plan(&self, spec: &LayaSpec) -> Graph

The forward pass as ops for a backend to lower, the graph in the module comment. Values are token rows until the markers are gathered, and a backend runs each op on the rows the batch has.

Trait Implementations

impl Clone for LayaGraph

fn clone(&self) -> Self

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for LayaGraph

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

Formats the value using the given formatter.

impl PartialEq for LayaGraph

fn eq(&self, other: &Self) -> bool

Equality operator ==.

fn ne(&self, other: &Rhs) -> bool

Inequality operator !=.

impl StructuralPartialEq for LayaGraph