Struct AprV2ReaderRef 

pub struct AprV2ReaderRef<'a> { /* private fields */ }

APR v2 format reader with zero-copy (borrows data - for mmap)

This reader borrows the data slice instead of copying it, enabling true zero-copy access when used with memory-mapped files.

Example

use apr_format::v2::AprV2ReaderRef;

let bytes: &[u8] = /* mmap'd slice */;
let reader = AprV2ReaderRef::from_bytes(bytes)?;
let weights = reader.get_f32_tensor("embed_tokens.weight")?;

Implementations

impl<'a> AprV2ReaderRef<'a>

pub fn from_bytes(data: &'a [u8]) -> Result<Self, V2FormatError>

Read from bytes (zero-copy - borrows data)

Unlike AprV2Reader::from_bytes, this does NOT copy the input data. The reader borrows the slice, making it ideal for use with mmap.

Errors

Returns error if parsing fails.

LAYOUT-002 Jidoka Guard

Rejects APR files with LAYOUT_COLUMN_MAJOR flag set, as these indicate improperly converted GGUF files that would produce garbage output.

pub fn header(&self) -> &AprV2Header

Get header

pub fn metadata(&self) -> &AprV2Metadata

Get metadata

pub fn tensor_names(&self) -> Vec<&str>

Get tensor names

pub fn get_tensor(&self, name: &str) -> Option<&TensorIndexEntry>

Get tensor by name

pub fn get_tensor_data(&self, name: &str) -> Option<&[u8]>

Get tensor data by name (zero-copy slice into mmap)

pub fn get_f32_tensor(&self, name: &str) -> Option<Vec<f32>>

Get tensor as f32 Vec (copies data from mmap to Vec<f32>)

Note: This allocates memory for the f32 values. For very large tensors, consider using get_tensor_data and processing in chunks.

pub fn verify_alignment(&self) -> bool

Check if all tensors are 64-byte aligned

pub fn tensor_index(&self) -> &[TensorIndexEntry]

Borrow the parsed tensor index (used by the core dequant extension).

Trait Implementations

impl<'a> Debug for AprV2ReaderRef<'a>

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

Formats the value using the given formatter.