Struct FlexTensor 

pub struct FlexTensor { /* private fields */ }

CPU tensor primitive for the Flex backend.

Uses type-erased byte storage with runtime dtype and Arc-based sharing. Clone is O(1) (refcount increment). Copy-on-write for mutations.

Implementations

impl FlexTensor

pub fn new(data: Bytes, layout: Layout, dtype: DType) -> FlexTensor

Create a new tensor from bytes, layout, and dtype.

pub fn from_data(data: TensorData) -> FlexTensor

Create a tensor from TensorData.

pub fn into_data(self) -> TensorData

Convert tensor to TensorData.

If non-contiguous or shared, this will copy data.

pub fn is_unique(&self) -> bool

Check if this tensor has exclusive ownership of its data.

When true, in-place mutations are safe without copying.

pub fn layout(&self) -> &Layout

Get the layout.

pub fn with_layout(self, layout: Layout) -> FlexTensor

Create a new tensor with a different layout but sharing the same data.

This is a zero-copy operation used for operations like flip, transpose, etc.

pub fn dtype(&self) -> DType

Get the dtype.

pub fn is_contiguous(&self) -> bool

Check if tensor is contiguous.

pub fn bytes(&self) -> &[u8]

Get the raw bytes (read-only).

pub fn data_arc(&self) -> Arc<Bytes>

Get a clone of the Arc for sharing data with a new layout.

Use this for zero-copy view operations (reshape, transpose, slice).

pub fn from_arc(data: Arc<Bytes>, layout: Layout, dtype: DType) -> FlexTensor

Create a tensor from shared data, layout, and dtype.

Use this for zero-copy view operations.

pub fn storage<E>(&self) -> &[E]

where E: Element + Pod,

Zero-copy typed view of the full storage buffer.

Use with StridedIter for non-contiguous access, or with layout().contiguous_offsets() for the contiguous fast path.

Panics

Panics if E::dtype() doesn’t match the tensor’s dtype. Note: Bool tensors are stored as u8, so both Bool(Native) and Bool(U8) dtypes accept u8 access.

pub fn storage_mut<E>(&mut self) -> &mut [E]

where E: Element + Pod,

Mutable typed view with copy-on-write semantics.

If the tensor is shared (refcount > 1), this will copy the data first. For in-place operations, prefer try_storage_mut() which returns None if shared, allowing you to choose an alternative strategy.

Panics

Panics if E::dtype() doesn’t match the tensor’s dtype. Note: Bool tensors are stored as u8, so both Bool(Native) and Bool(U8) dtypes accept u8 access.

pub fn try_storage_mut<E>(&mut self) -> Option<&mut [E]>

where E: Element + Pod,

Try to get mutable storage without copying.

Returns Some if tensor is uniquely owned, None if shared. Use this when you want to avoid the implicit copy in storage_mut(). Note: Bool tensors are stored as u8, so both Bool(Native) and Bool(U8) dtypes accept u8 access.

pub fn as_slice<E>(&self) -> Option<&[E]>

where E: Element + Pod,

Get typed slice view (zero-cost if contiguous and offset is 0).

Returns None if dtype doesn’t match E or tensor is non-contiguous.

pub fn empty(shape: Shape, dtype: DType) -> FlexTensor

Create an empty tensor with given shape and dtype.

pub fn zeros(shape: Shape, dtype: DType) -> FlexTensor

Create a tensor filled with zeros.

pub fn filled_typed<E>(shape: Shape, dtype: DType, value: E) -> FlexTensor

where E: Pod + Send + Sync,

Create a tensor filled with n copies of a typed value.

pub fn to_contiguous(&self) -> FlexTensor

Copy to contiguous layout if needed.

pub fn reshape(&self, new_shape: Shape) -> FlexTensor

Reshape tensor. Zero-copy if contiguous.

pub fn transpose(&self, dim1: usize, dim2: usize) -> FlexTensor

Transpose two dimensions. Zero-copy (metadata only).

pub fn narrow(&self, dim: usize, start: usize, len: usize) -> FlexTensor

Narrow/slice along a dimension. Zero-copy (metadata only).

pub fn permute(&self, axes: &[usize]) -> FlexTensor

Permute dimensions according to axes. Zero-copy (metadata only).

Trait Implementations

impl Clone for FlexTensor

fn clone(&self) -> FlexTensor

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for FlexTensor

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

Formats the value using the given formatter.

impl TensorMetadata for FlexTensor

type Device = FlexDevice

The device type associated with the tensor.

fn dtype(&self) -> DType

Get the dtype of the tensor.

fn shape(&self) -> Shape

Get the shape of the tensor.

fn rank(&self) -> usize

Get the number of dimensions of the tensor.

fn device(&self) -> <FlexTensor as TensorMetadata>::Device

Get the device associated with the tensor.

fn can_mut(&self) -> bool

Whether the tensor’s buffer can be mutated in place — i.e. this handle uniquely owns it, so an in-place op (slice_assign, an inplace kernel) writes the existing allocation instead of copying it first.

fn scheme(&self) -> QuantScheme

Get the quantization scheme for a quantized float tensor.