Struct DeepseekV4MtpModule 

pub struct DeepseekV4MtpModule {
    pub body: Box<dyn TransformerLayer>,
    pub enorm: DenseWeight,
    pub hnorm: DenseWeight,
    pub e_proj: DenseWeight,
    pub h_proj: DenseWeight,
    pub norm: DenseWeight,
    pub hc_head: Option<HcHeadWeights>,
}

2026-09-25: A loaded DeepSeek-V4 MTP draft module: the V4 layer body plus the MTP input combiner, final norm and HC head.

Fields

body: Box<dyn TransformerLayer>

2026-09-25: The V4 layer body, built by assemble_layer from the mtp.0 prefix.

enorm: DenseWeight

2026-09-25: RMSNorm weight applied to the token embedding before e_proj.

hnorm: DenseWeight

2026-09-25: RMSNorm weight applied to the target’s hidden state before h_proj.

e_proj: DenseWeight

2026-09-25: [hidden, hidden] projection of the normed embedding; summed with the h_proj branch.

h_proj: DenseWeight

2026-09-25: [hidden, hidden] projection of the normed hidden state; summed with the e_proj branch.

norm: DenseWeight

2026-09-25: Final RMSNorm weight, applied before the target’s LM head.

hc_head: Option<HcHeadWeights>

2026-09-25: The module’s own HC head (mtp.0.hc_head_*). The body is built with layer_idx = num_hidden_layers, so it is not the last model layer and never collapses the HC streams; the proposer does that with these weights. The body holds a clone of the same pointers. Some in every returned module: assemble_layer refuses hc_mult == 0.