Struct HyperHead 

pub struct HyperHead {
    pub fn_w: CudaSlice<f32>,
    pub base: CudaSlice<f32>,
    pub scale: CudaSlice<f32>,
}

Gated-head exit weights (HcCollapse::GatedHead, the dsv4 class). Absent under HcCollapse::Mean, which has no learned head (Glm5NextTextHyperHead is an unweighted mean).

Fields

fn_w: CudaSlice<f32>

base: CudaSlice<f32>

scale: CudaSlice<f32>

Implementations

impl HyperHead

pub fn load( e: &Engine, src: &dyn TensorSource, topology: &HyperTopology, hidden: usize, ) -> Result<Option<Self>, Box<dyn Error>>

None unless the collapse is gated. hc_head’s trio is shaped differently from a site’s: rows == streams and one scale (dsv4_forward::hc_head).