Struct Node 

pub struct Node {
    pub id: NodeId,
    pub name: String,
    pub op_type: String,
    pub domain: String,
    pub version: Option<i64>,
    pub inputs: Vec<Option<ValueId>>,
    pub outputs: Vec<ValueId>,
    pub attributes: HashMap<String, Attribute>,
    pub doc_string: Option<String>,
    pub device: Option<DeviceId>,
    pub exec_order: Option<usize>,
}

An operation in the graph.

Inputs are Option<ValueId> because ONNX ops may have optional (skipped) inputs represented by empty names; a None slot preserves positional arity. Outputs are always present (SSA values).

Fields

id: NodeId

name: String

Optional ONNX node name ("" means unnamed).

op_type: String

domain: String

Operator domain ("" == the default ONNX domain).

version: Option<i64>

Opset version of the operator called.

If None, the version is unspecified and follows the owning graph’s opset_imports. This property is special to ONNX IR to allow mixed opset usage in a graph for more flexible graph transformations; it does not exist in the ONNX protobuf spec. For example, a fusion may emit an opset-24 Swish into a graph whose other default-domain nodes still use the graph’s older exported opset, avoiding a false graph-wide upgrade.

inputs: Vec<Option<ValueId>>

outputs: Vec<ValueId>

attributes: HashMap<String, Attribute>

doc_string: Option<String>

device: Option<DeviceId>

Device placement, filled in by the placement pass.

exec_order: Option<usize>

Position in the final execution schedule, filled in by the scheduler.

Implementations

impl Node

pub fn new( id: NodeId, op_type: impl Into<String>, inputs: Vec<Option<ValueId>>, outputs: Vec<ValueId>, ) -> Self

A new node with the given op type and edges, and no attributes.

pub fn input_values(&self) -> impl Iterator<Item = ValueId> + '_

Iterate over the present (non-skipped) input value ids.

pub fn attr(&self, name: &str) -> Option<&Attribute>

Look up an attribute by name.

pub fn is_default_domain(&self) -> bool

Whether this node belongs to the default ONNX operator domain.

Relies on the post-load invariant that the loader canonicalizes the default domain to "" (see crate::normalize_domain), so this is a simple emptiness test — the "ai.onnx" spelling never reaches loaded IR.

pub fn local_opset(&self) -> Option<u64>

This node’s own opset version, if it names one that could be real.

The single owner of that judgement, so every subsystem reading Node::version agrees about the same node. Values that cannot be an opset — negative, zero, or beyond what any opset could plausibly reach — yield None: a node claiming them describes IR that is already wrong, and the graph’s own import is the better answer.

Callers with a graph in hand should prefer Graph::effective_opset, which falls back to that import. Shape inference has only the node and a map of imports, so it uses this directly.

Trait Implementations

impl Clone for Node

fn clone(&self) -> Node

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for Node

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

Formats the value using the given formatter.