forward_one_layer

metrale_model_weights::kimi_k3_host::cpu_forward

Function forward_one_layer 

Source 
pub fn forward_one_layer(
    ctx: &K3LayerCtx<'_>,
    layer: &K3CpuLayer,
    pos: usize,
    mixer_state: &mut LayerCache,
    stream: &mut AttnResStream,
    ablation: Ablation,
)
Expand description

One decoder layer: AttnRes + (KDA|MLA) + MLP. Mutates this layer’s LayerCache and the token’s AttnResStream.

Default cores are kda_decode_token / mla_decode_token / host mix_routed_experts. Serve CUDA injects via forward_one_layer_with_cores.