Struct UnifiedTensor 

pub struct UnifiedTensor {
    pub shape: Vec<usize>,
    pub dtype: DType,
    /* private fields */
}

A dense, row-major tensor stored in a GPU buffer.

Cloning is cheap: clones and reshaped views share the buffer.

Fields

shape: Vec<usize>

Dimensions, outermost first.

dtype: DType

Element type of the buffer.

Implementations

impl UnifiedTensor

pub fn buffer(&self) -> &Buffer

The underlying GPU buffer.

pub fn size_in_bytes(&self) -> u64

Size of the tensor’s data in bytes.

pub fn write_data(&self, context: &WgpuContext, data: &[u8])

Uploads data to the start of the buffer.

impl UnifiedTensor

pub fn new(context: &WgpuContext, data: &[f32], shape: &[usize]) -> Self

Uploads data as an f32 tensor of the given shape.

pub fn empty(context: &WgpuContext, shape: &[usize]) -> Self

An uninitialised f32 tensor drawn from the context’s buffer pool.

pub fn empty_with_dtype( context: &WgpuContext, shape: &[usize], dtype: DType, ) -> Self

An uninitialised tensor of dtype drawn from the buffer pool.

pub fn zeros(context: &WgpuContext, shape: &[usize]) -> Self

A zero-filled f32 tensor.

pub fn from_bytes(context: &WgpuContext, data: &[u8], shape: &[usize]) -> Self

Uploads raw bytes as an f32-typed tensor (e.g. packed quantised data).

pub fn from_bytes_with_dtype( context: &WgpuContext, data: &[u8], shape: &[usize], dtype: DType, ) -> Self

Uploads raw bytes as a tensor of dtype.

pub fn empty_fp16(context: &WgpuContext, shape: &[usize]) -> Self

Create an FP16 tensor (half precision for memory savings)

pub fn reshaped(&self, new_shape: &[usize]) -> Self

Create a new tensor view with different shape but sharing the same underlying data

pub async fn to_vec(&self, context: &WgpuContext) -> Vec<f32>

Copies the tensor back to the CPU, waiting for queued GPU work.

Trait Implementations

impl Clone for UnifiedTensor

fn clone(&self) -> Self

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for UnifiedTensor

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

Formats the value using the given formatter.