Struct InferenceContext 

pub struct InferenceContext<'a> {
    pub node: &'a Node,
    /* private fields */
}

The context passed to every op inference rule.

It exposes each input’s inferred type and shape-data, lets a rule mint fresh symbolic dimensions and broadcast shapes, and collects the outputs the rule produces. Rules never touch the Graph directly — they operate purely on this context, which makes them trivially unit testable in isolation.

Fields

node: &'a Node

The node being inferred.

Implementations

impl<'a> InferenceContext<'a>

pub fn new( node: &'a Node, inputs: Vec<NodeIo>, opset_imports: &'a HashMap<String, u64>, policy: MergePolicy, interner: &'a mut SymbolInterner, ) -> Self

Build a context for node from its resolved inputs (aligned with node.inputs, skipped slots carrying an empty NodeIo).

pub fn op(&self) -> &str

The op type of the node.

pub fn num_inputs(&self) -> usize

The number of input slots (including skipped optional ones).

pub fn num_outputs(&self) -> usize

The number of output slots.

pub fn has_input(&self, i: usize) -> bool

Whether input slot i is present (a value is connected).

pub fn input_type(&self, i: usize) -> Option<&TypeInfo>

The inferred type of input i, if resolved.

pub fn input_shape(&self, i: usize) -> Option<&[DimExpr]>

The inferred shape of input i, if resolved.

pub fn input_dtype(&self, i: usize) -> Option<DataType>

The inferred dtype of input i, if resolved.

pub fn input_rank(&self, i: usize) -> Option<usize>

The inferred rank of input i, if resolved.

pub fn input_shape_data(&self, i: usize) -> Option<&ShapeData>

The propagated shape-data of input i, if any.

pub fn set_output_type(&mut self, i: usize, type_info: TypeInfo)

Set the type of output i.

pub fn set_output(&mut self, i: usize, dtype: DataType, shape: TypedShape)

Set the dtype and shape of output i.

pub fn set_output_shape_data(&mut self, i: usize, data: ShapeData)

Set the propagated shape-data of output i.

pub fn into_outputs(self) -> Vec<NodeIo>

Consume the context, returning the outputs the rule produced.

pub fn policy(&self) -> MergePolicy

The active merge policy.

pub fn opset(&self, domain: &str) -> u64

The imported opset version for domain (the default ""/ai.onnx domain resolves to the canonical "" key, else 1).

pub fn fresh_dim(&mut self) -> DimExpr

Mint a fresh opaque dimension.

pub fn broadcast( &mut self, a: &[DimExpr], b: &[DimExpr], ) -> Result<TypedShape, ShapeInferError>

Broadcast two shapes under NumPy rules. Where two distinct symbolic dims must be unified, keeps a deterministic representative symbol (see broadcast_dim) rather than minting a fresh one. Errors only under MergePolicy::Strict on a concrete incompatibility.

pub fn broadcast_dim( &mut self, a: &DimExpr, b: &DimExpr, ) -> Result<DimExpr, ShapeInferError>

Broadcast a single pair of dimensions.