Function launch_dense_mlp 

pub fn launch_dense_mlp(
    layer: &K3BoundLayer,
    gpu: &dyn GpuBackend,
    x: &[f32],
    hidden: usize,
    inter: usize,
    beta: f32,
    linear_beta: f32,
    stream: u64,
) -> Result<Vec<f32>>

Dense and shared expert matrices are already TP-sharded in the loader.