Struct GradNode 

pub struct GradNode<Op: GraphOp> { /* private fields */ }

Backward computation node for eager reverse-mode AD.

A GradNode records one primal operation, the stable input aliases used to replay that operation during backward, the user-visible output keys that can receive cotangent seeds, and the edges to parent eager values.

Examples

use tidu::{GradEdge, GradNode};

let node = GradNode::new(
    op,
    input_aliases,
    output_keys,
    saved_forward_values,
    vec![GradEdge::new(parent_node, input_key, true)],
);

Implementations

impl<Op: GraphOp> GradNode<Op>

pub fn new( op: Op, primal_in_keys: Vec<GlobalValKey<Op>>, primal_out_keys: Vec<GlobalValKey<Op>>, saved_data: HashMap<GlobalValKey<Op>, Arc<Op::Operand>>, input_edges: Vec<GradEdge<Op>>, ) -> Self

Create a grad node and validate the shape of its eager AD metadata.

primal_in_keys must contain GlobalValKey::Input aliases. The eager backward path linearizes one operation at a time and rebuilds those aliases as fragment inputs.

Examples

let node = tidu::GradNode::new(
    op,
    input_aliases,
    output_keys,
    saved_data,
    input_edges,
);

pub fn op(&self) -> &Op

The primal operation recorded by this node.

pub fn primal_in_keys(&self) -> &[GlobalValKey<Op>]

Stable input aliases used for single-op backward replay.

pub fn primal_out_keys(&self) -> &[GlobalValKey<Op>]

User-visible output keys, one per primal output slot.

pub fn saved_data(&self) -> &HashMap<GlobalValKey<Op>, Arc<Op::Operand>>

Saved concrete primal input and derived output values.

pub fn input_edges(&self) -> &[GradEdge<Op>]

Edges to the eager values that provided this node’s inputs.