bind_kda_weights

spark_model::layers::glm5next_kda::binding

Function bind_kda_weights 

Source 
pub fn bind_kda_weights(
    gpu: &dyn GpuBackend,
    cfg: &Glm5NextKdaConfig,
    layer_idx: usize,
    src: &dyn KdaTensorSource,
) -> Result<(Glm5NextKdaWeights, KdaBindReport)>
Expand description

Bind one KDA block. Strict: exact tensor set, exact dtypes, exact shapes.

Weights are uploaded verbatim — BF16 stays BF16, F32 stays F32, nothing is converted, requantised or dequantised, because nothing in a KDA block is quantised in the first place.