Struct AprV2Writer 

pub struct AprV2Writer { /* private fields */ }

APR v2 format writer

Implementations

impl AprV2Writer

pub fn new(metadata: AprV2Metadata) -> Self

Create new writer

LAYOUT-002: All new APR files are created with LAYOUT_ROW_MAJOR flag set. This ensures realizar can safely assume row-major layout for all tensors.

pub fn add_tensor( &mut self, name: impl Into<String>, dtype: TensorDType, shape: Vec<usize>, data: Vec<u8>, )

Add tensor to the file

pub fn add_f32_tensor( &mut self, name: impl Into<String>, shape: Vec<usize>, data: &[f32], )

Add f32 tensor

pub fn add_f16_tensor( &mut self, name: impl Into<String>, shape: Vec<usize>, data: &[f32], )

Add f16 tensor (converts f32 → f16, 2 bytes per value)

This provides true 2x compression over f32 storage with minimal precision loss for inference workloads. Uses IEEE 754 half-precision format.

pub fn add_q8_tensor( &mut self, name: impl Into<String>, shape: Vec<usize>, data: &[f32], )

Add Q8 tensor (8-bit symmetric quantization)

Format: [scale: f32 (4 bytes)] + [quantized: i8 × n] Total size: 4 + n bytes (vs 4n for f32) Compression ratio: ~4x

pub fn add_q4_tensor( &mut self, name: impl Into<String>, shape: Vec<usize>, data: &[f32], )

Add Q4 tensor (4-bit symmetric quantization, block-wise)

Format: For each block of 32 values: [block_scale: f16 (2 bytes)] + [packed nibbles: 16 bytes]

Total size per block: 18 bytes (vs 128 bytes for f32) Compression ratio: ~7x

pub fn add_q4k_raw_tensor( &mut self, name: impl Into<String>, shape: Vec<usize>, raw_data: Vec<u8>, )

Add raw Q4_K tensor (GGUF-compatible super-block format)

This stores GGUF Q4_K data directly without re-quantization. Q4_K format: 256-element super-blocks with nested 32-element sub-blocks Each super-block: d (f16, 2B) + dmin (f16, 2B) + scales (12B) + qs (128B) = 144 bytes Effective bits per weight: ~4.5

Use this when importing from GGUF to preserve exact quantization.

pub fn add_q6k_raw_tensor( &mut self, name: impl Into<String>, shape: Vec<usize>, raw_data: Vec<u8>, )

Add raw Q6_K tensor (GGUF-compatible super-block format)

This stores GGUF Q6_K data directly without re-quantization. Q6_K format: 256-element super-blocks Each super-block: ql (128B) + qh (64B) + scales (16B) + d (f16, 2B) = 210 bytes Effective bits per weight: ~6.5

pub fn with_lz4_compression(&mut self) -> &mut Self

Set LZ4 compression flag

pub fn with_sharding( &mut self, shard_count: usize, shard_index: usize, ) -> &mut Self

Set sharding info

pub fn write(&mut self) -> Result<Vec<u8>, V2FormatError>

Write to bytes

Errors

Returns error if serialization fails.

pub fn write_to<W: Write>( &mut self, writer: &mut W, ) -> Result<(), V2FormatError>

Write to a Write impl

Errors

Returns error if write fails.

Trait Implementations

impl Debug for AprV2Writer

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

Formats the value using the given formatter.