Struct Tensor 

pub struct Tensor { /* private fields */ }

A multi-dimensional tensor with reference-counted buffer ownership.

Implementations

impl Tensor

pub fn zeros(shape: impl Into<Shape>, dtype: DType) -> Result<Self>

Create a zero-filled tensor on the CPU.

pub fn from_f32(data: &[f32], shape: impl Into<Shape>) -> Result<Self>

Create a tensor from a flat f32 slice (CPU, row-major).

pub fn from_bf16_bytes(data: &[u8], shape: impl Into<Shape>) -> Result<Self>

Create a tensor from raw BF16 bytes, storing them natively without conversion. Use to_f32_vec() or to_f32_tensor() to convert for computation.

pub fn from_f16_bytes(data: &[u8], shape: impl Into<Shape>) -> Result<Self>

Create a tensor from raw F16 bytes, storing them natively without conversion.

pub fn from_quant_bytes( data: &[u8], shape: impl Into<Shape>, dtype: DType, ) -> Result<Self>

Create a quantized tensor from raw block bytes (Q4_0 / Q8_0).

data must contain exactly dtype.byte_count(shape.numel()) bytes, i.e. the packed ggml block bytes with no expansion. The shape describes the logical element count; shape.numel() must be a multiple of 32.

pub fn scalar_f32(v: f32) -> Result<Self>

Create a scalar tensor from a single f32.

pub fn from_buffer( shape: impl Into<Shape>, dtype: DType, buffer: BufferHandle, offset: usize, ) -> Result<Self>

Create from a pre-built BufferHandle (used by backends).

pub fn shape(&self) -> &Shape

pub fn dtype(&self) -> DType

pub fn ndim(&self) -> usize

pub fn numel(&self) -> usize

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

pub fn buffer(&self) -> &BufferHandle

pub fn offset(&self) -> usize

pub fn is_scalar(&self) -> bool

True if the tensor has a single element.

pub fn is_contiguous(&self) -> bool

True if the buffer is row-major contiguous (normal case).

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

Raw byte view. For non-quantized tensors returns the full buffer slice from offset onwards (preserving the original behavior that stride-based kernels rely on). For quantized tensors (Q4_0/Q8_0) returns exactly the packed block bytes for this tensor’s logical shape.

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

For quantized tensors (Q4_0, Q8_0): returns the packed block bytes as a row-major slice where each logical row of k elements occupies dtype.byte_count(k) bytes. Panics if the tensor is not quantized.

pub fn as_f32_slice(&self) -> &[f32]

Typed f32 view — panics if dtype is not F32.

pub fn to_f32_cow(&self) -> Cow<'_, [f32]>

Returns a Cow<[f32]>. Borrows if the tensor is already F32, otherwise allocates a new Vec<f32>.

pub fn to_f32_vec(&self) -> Vec<f32>

Convert this tensor to a Vec<f32>, handling all dtypes including quantized. For F32: cheap copy. For F16/BF16: convert. For Q4_0/Q8_0: dequantize all blocks.

pub fn to_f32_tensor(&self) -> Result<Tensor>

Returns an F32 tensor, converting BF16/F16 if necessary. For already-F32 tensors, clones the buffer. For native types, converts.

pub fn as_f32_slice_mut(&mut self) -> Result<&mut [f32]>

Mutable typed f32 view — fails if buffer is shared or not F32.

pub fn reshape(&self, new_shape: impl Into<Shape>) -> Result<Tensor>

Returns a new tensor with a different shape but the same buffer. The total number of elements must be unchanged.

pub fn t(&self) -> Result<Tensor>

Transpose a 2-D tensor (swap axes 0 and 1).

pub fn slice_axis( &self, axis: usize, start: usize, end: usize, ) -> Result<Tensor>

Return a view of the tensor sliced along the given axis.

pub fn byte_size(&self) -> usize

Byte count for all elements.

Trait Implementations

impl Clone for Tensor

fn clone(&self) -> Tensor

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for Tensor

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

Formats the value using the given formatter.

impl<'de> Deserialize<'de> for Tensor

fn deserialize<D: Deserializer<'de>>(deserializer: D) -> Result<Self, D::Error>

Deserialize this value from the given Serde deserializer.

impl Display for Tensor

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

Formats the value using the given formatter.

impl From<&Tensor> for TensorMeta

fn from(t: &Tensor) -> Self

Converts to this type from the input type.

impl Serialize for Tensor

fn serialize<S: Serializer>(&self, serializer: S) -> Result<S::Ok, S::Error>

Serialize this value into the given Serde serializer.