Skip to main content

expand

cubecl_std::quant::dequantize_aligned

Function expand 

Source 
pub fn expand<Q: Scalar, S: CubePrimitive, F: Numeric, NQ: Size, NF: Size>(
    scope: &Scope,
    value: <Vector<Q, NQ> as CubeType>::ExpandType,
    scale: <S as CubeType>::ExpandType,
    table: <ComptimeOption<Box<[f32]>> as CubeType>::ExpandType,
    scheme: QuantScheme,
) -> <Vector<F, NF> as CubeType>::ExpandType
Expand description

Dequantize a vector of values, where vector_size * num_quants is a power of two. Unaligned values can’t be dequantized in place.

scale is the effective scale for these values: how many scale levels the scheme has and how they combine is the caller’s business, folded before the call. This is what keeps the primitive per-read arithmetic, serving equally a one-level read, a view multiplying the per-tensor scale in per read, or a tile handing every read one register. table is a lookup scheme’s 2^bits-entry table, and must be present exactly when scheme.mode is QuantMode::Lookup.