Struct Nf4Quantizer 

pub struct Nf4Quantizer {
    pub block_size: usize,
}

NF4 quantizer — encodes tensors to packed 4-bit NF4 codes.

Blocks of block_size elements share an absmax scaling factor. The default block_size of 64 matches the QLoRA paper.

Fields

block_size: usize

Number of elements per absmax scaling block.

Implementations

impl Nf4Quantizer

pub fn new(block_size: usize) -> Self

Create an NF4 quantizer with the given block size.

Panics

Panics if block_size == 0.

pub fn encode(&self, tensor: &[f32]) -> QuantResult<(Vec<u8>, Vec<f32>)>

Encode a flat tensor to packed NF4 bytes and per-block absmax values.

The number of elements must be a multiple of block_size, and block_size must be even (pairs of nibbles pack into bytes).

Returns (packed_bytes, absmax_per_block).

Errors

• QuantError::GroupSizeMismatch — if len is not divisible by block_size.
• QuantError::EmptyInput — if tensor is empty.

pub fn decode(&self, packed: &[u8], absmaxs: &[f32]) -> QuantResult<Vec<f32>>

Decode packed NF4 bytes back to f32 using the stored absmax values.

Errors

• QuantError::DimensionMismatch — if packed/absmax lengths are inconsistent.

pub fn quantization_mse(&self, tensor: &[f32]) -> QuantResult<f32>

Estimate the quantization error (mean squared error) for a tensor.

Errors

Propagates errors from encode / decode.

Trait Implementations

impl Clone for Nf4Quantizer

fn clone(&self) -> Nf4Quantizer

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for Nf4Quantizer

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

Formats the value using the given formatter.

impl Default for Nf4Quantizer

fn default() -> Self

Returns the “default value” for a type.