Enum Glm5NextMixer 

pub enum Glm5NextMixer {
    Kda {
        layer: Box<Glm5NextKdaLayer>,
        ws: Arc<Glm5NextKdaWorkspace>,
        cfg: Glm5NextKdaConfig,
    },
    Dsa(Box<Glm5NextDsaLayer>),
}

Which mixer this layer runs. Both halves already exist and are GPU-gated; this enum is the dispatch, not new math.

🪤 KDA’s decode/prefill are inherent methods with their own signature, not the TransformerLayer ones — they take a &KdaSeqState and a workspace and leave the result in ws.final_out. DSA’s decode is the trait method and writes its result back into the buffer it was handed. The two conventions differ; this is where they are reconciled.

Variants

Kda

Fields

layer: Box<Glm5NextKdaLayer>

ws: Arc<Glm5NextKdaWorkspace>

Shared across every KDA layer — all 34 have identical geometry, so one workspace serves them all. Arc because the layers are independent owners.

cfg: Glm5NextKdaConfig

Dsa(Box<Glm5NextDsaLayer>)