Struct ScaleBindings 

pub struct ScaleBindings<C: Coordinates, R: Runtime> { /* private fields */ }

The scale bindings of a quantized view, one per scheme level.

Only the block scales are addressed per position, so they bind as a view; the per-tensor scale of a two-level scheme covers the whole tensor and binds as a buffer holding its one scale in the first element, read once per kernel as f32.

Implementations

impl<C: Coordinates, R: Runtime> ScaleBindings<C, R>

pub fn one(scales: ViewArg<C, R>) -> Self

The binding of a one-level scheme’s scales.

pub fn two(scales: ViewArg<C, R>, global_scale: BufferArg<R>) -> Self

The bindings of a two-level scheme: the block scales and the per-tensor scale they are normalized against.

pub fn lookup(scales: ViewArg<C, R>, table: BufferArg<R>) -> Self

The bindings of a one-level lookup scheme (QuantMode::Lookup): the scales and the table each stored field indexes, so a read reconstructs table[field] * scale.

table must hold 2^bits f32 entries — registration checks its length against the scheme, and the unpack’s mask bounds every index to that range.

pub fn len(&self) -> usize

The number of bound scale levels, matched against the scheme’s at construction: the inner binding plus the global scale when bound. The lookup table is not a level and is never counted; its presence is checked against the scheme’s mode instead (quant::check_table_bindings).