Struct TensorReader 

pub struct TensorReader<'a> { /* private fields */ }

A lazy view over one tensor’s raw bytes inside a ModelSource.

The byte slice borrows from the source (e.g. an mmap region) — no copy is made until TensorReader::load_data is called. Format adapters that decode on open (e.g. GGUF quantization) use TensorReader::owned.

Implementations

impl<'a> TensorReader<'a>

pub fn new( name: String, shape: Vec<usize>, dtype: TensorDtype, data: &'a [u8], ) -> Self

Creates a reader from raw parts. data must be shape.iter().product::<usize>() * dtype.size() little-endian bytes.

pub fn owned(name: String, shape: Vec<usize>, data: Vec<u8>) -> Self

Creates a reader over owned (already-decoded) f32 bytes.

pub fn owned_with_dtype( name: String, shape: Vec<usize>, dtype: TensorDtype, data: Vec<u8>, ) -> Self

Creates a reader over owned bytes of an explicit dtype (used when a passthrough tensor had to be row-reordered on load).

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

Element shape.

pub fn dtype(&self) -> TensorDtype

On-disk dtype.

pub fn num_elements(&self) -> usize

Number of elements.

pub fn load_data(&self) -> Result<TensorData>

Converts the raw bytes to f32 TensorData. Integer and boolean tensors are cast to f32 so weight loaders never abort on buffer dtypes such as I64 position_ids.

pub fn load_to_tensor<B: Backend, const D: usize>( &self, device: &Device<B>, ) -> Result<Tensor<B, D>>

Loads the tensor onto a backend device as an f32 tensor of rank D.

pub fn load_int_tensor<B: Backend, const D: usize>( &self, device: &Device<B>, ) -> Result<Tensor<B, D, Int>>

Loads raw unsigned-byte data onto a backend device as an i32 tensor of rank D (values 0..=255). Only valid for U8 tensors; used to feed packed quantized weights to GPU-side dequantization.

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

The raw little-endian bytes, borrowing from the source (zero-copy).