Struct HcWeights 

pub struct HcWeights {
    pub attn: HcSiteWeights,
    pub ffn: HcSiteWeights,
    pub head: Option<HcHeadWeights>,
    pub hc_mult: usize,
    pub sinkhorn_iters: usize,
    pub hc_eps: f32,
    pub is_first_model_layer: bool,
    pub is_last_model_layer: bool,
}

Fields

attn: HcSiteWeights

ffn: HcSiteWeights

head: Option<HcHeadWeights>

Model-level head weights. Some on all layers (replicated pointer), consumed only by the last layer’s hc_head call.

hc_mult: usize

sinkhorn_iters: usize

hc_eps: f32

is_first_model_layer: bool

Whether this is model layer 0 — the layer that seeds the highway with hc_expand.

Carried here rather than derived from attn_layer_idx, which counts ATTENTION layers. On DeepSeek-V4 every layer is attention and the two indices coincide; on a 3:1 GDN:attention interleave they do not, and attn_layer_idx == 0 is model layer 3 — three layers after the highway needed seeding.

is_last_model_layer: bool

Whether this is the LAST model layer — the one that collapses the highway with hc_head.

Same reason. The old guard was attn_layer_idx + 1 == num_hidden_layers, i.e. 12 == 48 on this model: never true, so the LM head would have read an uncollapsed stream. On Qwen that also means an UNNORMALIZED one, since hyper_connection_mixer is the model’s final norm and the checkpoint ships no model.norm.weight.