Skip to main content

multiply_global_scale

cubecl_std::quant

Function multiply_global_scale 

Source 
pub fn multiply_global_scale<S: CubePrimitive>(
    global_scale: f32,
    scale: S,
) -> f32
Expand description

The effective scale of values whose per-tensor scale multiplies on top of their block scale.

The two multiply in f32: a block scale is normalized against the per-tensor one, so on its own it overflows a narrow compute type by orders of magnitude before the global scale can bring the product back into range.