Function hc_head_site 

pub fn hc_head_site(
    gpu: &dyn GpuBackend,
    kernel: KernelHandle,
    streams: DevicePtr,
    head: &HcHeadWeights,
    hc: &HcWeights,
    y_out: DevicePtr,
    scratch: DevicePtr,
    num_tokens: u32,
    hidden_size: u32,
    norm_eps: f32,
    stream: u64,
) -> Result<()>

The model-level final collapse before the LM head.

On Qwen this is ALSO the model’s final normalization — the checkpoint ships no model.norm.weight because hyper_connection_mixer’s hc_norm plays that role.