Crate tidu 

AD graph transforms for the tensor4all v2 stack.

This crate provides two graph-to-graph transforms: differentiate for forward linearization (JVP) and transpose for reverse linear flow over a linear fragment. It also provides eager reverse-mode AD helpers: record_eager_op builds GradNode metadata around concrete frontend execution, and backward_dag replays recorded nodes through caller-provided BackwardCallbacks.

Examples

use computegraph::resolve::resolve;
use tidu::{differentiate, transpose};

let view = resolve(vec![primal_fragment]);
let mut ctx = ();
let linear = differentiate(&view, &[output_key], &[input_key], 1, &mut ctx);
let _transposed = transpose(&linear, &mut ctx);

Re-exports

pub use backward::backward_dag;

pub use backward::topo_sort_grad_dag;

pub use backward::BackwardCallbacks;

pub use eager_transpose::eager_transpose_fragment;

pub use grad_node::GradEdge;

pub use grad_node::GradNode;

Modules

backward

eager_transpose

grad_node

Structs

EagerOutput

Per-output trace metadata returned by record_eager_op.

EagerValue

Eager frontend input descriptor for generic AD recording.

LinearFragment

A linear fragment produced by crate::differentiate or crate::transpose.

Traits

EagerKeySource

Caller-provided source of stable eager value keys.

Functions

derived_output_key

Construct the derived key used to save a replayed primal output value.

differentiate

Differentiate a resolved computation graph, producing a linear fragment.

record_eager_op

Record a concrete eager primitive execution for reverse-mode AD.

saved_forward_values

Build saved forward data for one eager op.

transpose

Transpose a linear fragment, reversing linear flow.