Struct WeightsConfig 

pub struct WeightsConfig {
    pub num_bits: u8,
    pub type_: String,
    pub symmetric: bool,
    pub strategy: String,
    pub axis: Option<usize>,
    pub group_size: Option<usize>,
    pub fp8_variant: Option<String>,
}

The compressed-tensors weights block for one scheme.

Fields

num_bits: u8

Bits per quantized element.

type_: String

"int" or "float".

symmetric: bool

Whether zero_point is implicitly 0 (no _zero_point tensor is written).

strategy: String

"tensor", "channel", or "group" (see Granularity::strategy_name).

axis: Option<usize>

The axis grouping runs along, present for "channel"/"group" strategies. Required to reconstruct the exact element->group mapping (see crate::quant::compute_groups) – strategy and group_size alone are ambiguous for tensors with rank > 2.

group_size: Option<usize>

Elements per group, only present when strategy == "group".

fp8_variant: Option<String>

"e4m3" or "e5m2", present only for type == "float". The two FP8 encodings are not bit-compatible with each other, so this is required to correctly decode – without it, crate::validate would have to guess the variant used to write the file.

Trait Implementations

impl Clone for WeightsConfig

fn clone(&self) -> WeightsConfig

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for WeightsConfig

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

Formats the value using the given formatter.

impl<'de> Deserialize<'de> for WeightsConfig

fn deserialize<__D>(__deserializer: __D) -> Result<Self, __D::Error>

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl Serialize for WeightsConfig

fn serialize<__S>(&self, __serializer: __S) -> Result<__S::Ok, __S::Error>

where __S: Serializer,

Serialize this value into the given Serde serializer.