Struct TensorGraph 

pub struct TensorGraph {
    pub identity: Option<Identity>,
    pub io: Option<StepIo>,
    pub tensors: BTreeMap<String, Tensor>,
    pub ops: Vec<Op>,
}

A cut of a model’s one program as a graph.

Fields

identity: Option<Identity>

The cut it was lifted from.

io: Option<StepIo>

The names it reads and writes across steps.

tensors: BTreeMap<String, Tensor>

ops: Vec<Op>

The ops in program order, the schedule order every backend starts from.

Implementations

impl TensorGraph

pub fn lift( rp: &RecognizedProgram, identity: Option<Identity>, io: Option<StepIo>, storage: &dyn Fn(&str) -> Storage, ) -> Result<TensorGraph, GraphError>

The graph of a walked program: storage says where each constant’s values live (the IR reads no checkpoint), and identity and io are the cut it was lifted from and the names it carries across steps.

pub fn placed( &self, storage: &dyn Fn(&str) -> Storage, io: Option<StepIo>, ) -> TensorGraph

The graph with each constant’s values where storage says and io the names it carries across steps, which its host reads: a graph lifted before its constants were placed, placed as TensorGraph::lift places them.

pub fn lift_stages( rp: &RecognizedProgram, identity: Option<Identity>, io: Option<StepIo>, storage: &dyn Fn(&str) -> Storage, ) -> TensorGraph

The graph of the statements the front end lifted, whatever else the program holds: a backend that reports the statements it could not lift on its own lowers the rest. A tensor a statement the front end passed over computes is an activation of the shape the program gives it.

pub fn lift_source( source: &str, storage: &dyn Fn(&str) -> Storage, ) -> Result<Lifted, SpatialReject>

The front end’s walk of source and the graph of what it lifts (TensorGraph::lift_stages): the one way a backend reaches the ops of a source.

pub fn of_cut( spec: &ModelSpec, cut: Cut, io: Option<StepIo>, storage: &dyn Fn(&str) -> Storage, ) -> Result<TensorGraph, GraphError>

The graph of the cut cut of spec’s program, recording the cut’s identity: io the names it carries across steps, storage where each constant’s values live.

pub fn of_op( op: TensorOpSpec, ins: Vec<(Symbol, usize, Fill)>, out: Symbol, interner: &Interner, ) -> TensorGraph

The graph of one op over operands the source declares, each (symbol, length, fill), and its result out: a statement a backend recognizes on its own, outside any walk of a program. Every symbol the op’s spec names is one of its operands or its result.

pub fn shape_of(&self, tensor: &str, n: usize) -> (usize, usize)

The (rows, cols) the program gives tensor, where that accounts for the n elements an op counts — the 1 × n row otherwise.

pub fn touches(&self) -> Touches<'_>

Which ops touch each tensor — read it, write it, or name the rows it lands in — taken once, for questions asked of many tensors.

pub fn of_role(&self, role: Role) -> impl Iterator<Item = &Tensor>

The tensors of one role, by name.

pub fn readers(&self, tensor: &str) -> Vec<usize>

The ops that read tensor, by index.

pub fn producer(&self, tensor: &str) -> Option<usize>

The op that computes tensor into its own buffer.

pub fn writes_in_place(&self) -> BTreeSet<String>

Every tensor an op writes in place, however deep in a fused op it is written: what it updates as state, and every table or wider tensor a destination lands rows or units in.

pub fn caches(&self) -> BTreeSet<String>

The attention caches: every state tensor an attention reads its keys or values from, itself or through the head slices cut from it (a latent cache’s values are its rows’ first part).

Trait Implementations

impl Clone for TensorGraph

fn clone(&self) -> TensorGraph

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for TensorGraph

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

Formats the value using the given formatter.