mix_routed_experts

atlas_core::kimi_k3::latent_moe

Function mix_routed_experts 

Source 
pub fn mix_routed_experts(
    latent: &[f32],
    ids: &[usize],
    weights: &[f32],
    experts: &[(Vec<f32>, Vec<f32>, Vec<f32>)],
    cfg: &LatentMoeConfig,
) -> Vec<f32>
Expand description

Mix selected SiTU-GLU experts into latent. Empty w1 means packed-only (no host dequant) — CUDA grouped GEMM must have run instead.