Struct Shape 

pub struct Shape { /* private fields */ }

Tensor shape: ordered list of dimensions + element type.

SmallVec<[Dim; 4]> avoids heap allocation for up to 4D tensors (the common case).

Implementations

impl Shape

pub fn batch_seq(batch: u32, seq: u32, hidden: usize, dtype: DType) -> Self

[batch, seq, hidden] with symbolic leading axes.

pub fn batch_seq_2d(batch: u32, seq: u32, dtype: DType) -> Self

[batch, seq] matrix.

pub fn batch_seq_heads( batch: u32, seq: u32, heads: usize, head_dim: usize, dtype: DType, ) -> Self

[batch, seq, heads, head_dim] attention layout.

impl Shape

pub fn new(dims: &[usize], dtype: DType) -> Self

Create a shape from static dimensions.

pub fn from_dims(dims: &[Dim], dtype: DType) -> Self

Create a shape with mixed static/dynamic dimensions.

pub fn scalar(dtype: DType) -> Self

Scalar (0-dimensional).

pub fn rank(&self) -> usize

pub fn dtype(&self) -> DType

pub fn dims(&self) -> &[Dim]

pub fn dim(&self, i: usize) -> Dim

pub fn gather_index_count(&self, axis: usize) -> usize

How many positions a gather/scatter along axis takes, when self is the shape of the INDEX operand.

Gather’s output rank is data_rank - 1 + index_rank, so an index may be either:

• rank > axis — the usual case, where the index carries a dimension at the gather axis and that dimension is the count; or
• a flat list (typically rank 1) that has no dimension at axis and whose whole extent IS the count.

Every backend previously wrote idx_shape.dim(axis) here, which panics with “Shape::dim(1) out of bounds for rank 1” on the second form — so a rank-1 gather built a correct forward graph and then blew up the moment it needed a backward pass. Sharing the rule keeps the five backends that need it from disagreeing about it.

pub fn dynamic_symbols(&self) -> Vec<u32>

Set of dynamic dim symbols this shape references. Useful for “what bindings does this graph need?” queries on inputs.

pub fn bind(&self, bindings: &DimBinding) -> Self

Specialize the shape against a binding (symbol → static size). Unknown symbols stay Dim::Dynamic. Plan #54: the step that takes a “compile once, run at any seq length” graph and produces the runtime-specific concrete shape.

pub fn num_elements(&self) -> Option<usize>

Total number of elements (only if all dims are static).

pub fn size_bytes(&self) -> Option<usize>

Total size in bytes (only if all dims are static).

pub fn is_static(&self) -> bool

True if all dimensions are statically known.

pub fn with_dim(self, axis: usize, dim: Dim) -> Self

Replace a dimension.

pub fn with_dtype(self, dtype: DType) -> Self

Change dtype (for cast operations).

pub fn broadcast_with(&self, other: &Shape) -> Result<Shape, String>

Numpy-style broadcast with another shape (fusion / lowering).

Trait Implementations

impl Clone for Shape

fn clone(&self) -> Shape

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for Shape

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

Formats the value using the given formatter.

impl<'de> Deserialize<'de> for Shape

fn deserialize<__D>(__deserializer: __D) -> Result<Self, __D::Error>

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl Display for Shape

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

Formats the value using the given formatter.

impl Eq for Shape

impl Hash for Shape

fn hash<__H: Hasher>(&self, state: &mut __H)

Feeds this value into the given Hasher.

fn hash_slice<H>(data: &[Self], state: &mut H)

where H: Hasher, Self: Sized,

Feeds a slice of this type into the given Hasher.

impl PartialEq for Shape

fn eq(&self, other: &Shape) -> bool

Equality operator ==.

fn ne(&self, other: &Rhs) -> bool

Inequality operator !=.

impl Serialize for Shape

fn serialize<__S>(&self, __serializer: __S) -> Result<__S::Ok, __S::Error>

where __S: Serializer,

Serialize this value into the given Serde serializer.

impl StructuralPartialEq for Shape