Struct CudaTensor 

pub struct CudaTensor { /* private fields */ }

Tensor stored on CUDA device

Implementations

impl CudaTensor

pub fn new(shape: &[usize], dtype: DType, device_id: i32) -> CudaResult<Self>

Create new CUDA tensor with given shape (uninitialized)

pub fn from_tensor(tensor: &Tensor, device_id: i32) -> CudaResult<Self>

Create CUDA tensor from CPU tensor (copies data to GPU)

pub fn from_tensor_async( tensor: &Tensor, device_id: i32, stream: &CudaStream, ) -> CudaResult<Self>

Create CUDA tensor from CPU tensor (async copy)

pub fn to_tensor(&self) -> CudaResult<Tensor>

Copy back to CPU tensor

pub fn to_tensor_async(&self, stream: &CudaStream) -> CudaResult<Vec<f32>>

Copy back to CPU tensor (async)

pub fn shape(&self) -> &Shape

Get shape

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

Get dimensions

pub fn numel(&self) -> usize

Get number of elements

pub fn dtype(&self) -> DType

Get data type

pub fn device_id(&self) -> i32

Get device ID

pub fn as_ptr(&self) -> *const f32

Get raw buffer pointer

pub fn as_mut_ptr(&mut self) -> *mut f32

Get mutable raw buffer pointer

pub fn size_bytes(&self) -> usize

Get buffer size in bytes

pub fn zeros(shape: &[usize], device_id: i32) -> CudaResult<Self>

Create zeros tensor

pub fn ones(shape: &[usize], device_id: i32) -> CudaResult<Self>

Create ones tensor

pub fn full(shape: &[usize], value: f32, device_id: i32) -> CudaResult<Self>

Create tensor filled with value

pub fn reshape(&self, new_shape: &[usize]) -> CudaResult<Self>

Reshape tensor

pub fn transpose(&self, dim0: usize, dim1: usize) -> CudaResult<Self>

Transpose dimensions

pub fn is_contiguous(&self) -> bool

Check if tensor is contiguous

pub fn clone_tensor(&self) -> CudaResult<Self>

Deep clone

pub fn add(&self, other: &CudaTensor) -> CudaResult<CudaTensor>

Element-wise addition

pub fn sub(&self, other: &CudaTensor) -> CudaResult<CudaTensor>

Element-wise subtraction

pub fn mul_scalar(&self, scalar: f32) -> CudaResult<CudaTensor>

Scalar multiplication

pub fn matmul(&self, other: &CudaTensor) -> CudaResult<CudaTensor>

Matrix multiplication

pub fn dot(&self, other: &CudaTensor) -> CudaResult<f32>

Dot product

pub fn norm(&self) -> CudaResult<f32>

L2 norm

pub fn sum(&self) -> CudaResult<f32>

Sum all elements

pub fn mean(&self) -> CudaResult<f32>

Mean of all elements

pub fn max(&self) -> CudaResult<f32>

Max element

pub fn min(&self) -> CudaResult<f32>

Min element

pub fn relu(&self) -> CudaResult<CudaTensor>

ReLU activation (in-place capable)

pub fn sigmoid(&self) -> CudaResult<CudaTensor>

Sigmoid activation

pub fn gelu(&self) -> CudaResult<CudaTensor>

GELU activation

pub fn softmax(&self, dim: i32) -> CudaResult<CudaTensor>

Softmax

pub fn exp(&self) -> CudaResult<CudaTensor>

Exponential

pub fn log(&self) -> CudaResult<CudaTensor>

Natural logarithm

pub fn sqrt(&self) -> CudaResult<CudaTensor>

Square root

pub fn pow(&self, exp: f32) -> CudaResult<CudaTensor>

Power

Trait Implementations

impl Clone for CudaTensor

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 CudaTensor

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

Formats the value using the given formatter.