Function forward_token_graph 

pub fn forward_token_graph(
    model: &Arc<CmfModel>,
    kv_id: u64,
    layers: &[GraphLayer<'_>],
    invf: &[f32],
    h: &mut [f32],
    nh: usize,
    nkv: usize,
    hd: usize,
    rd: usize,
    hidden: usize,
    inter: usize,
    position: usize,
    cap: usize,
    gemma: bool,
    eps: f32,
    lm_head: Option<(&GraphW<'_>, usize)>,
    final_norm: &[f32],
    logits: &mut Vec<f32>,
    loop_norm_at: &[usize],
) -> bool

Whole-token decode graph on wgpu: the entire layer stack in ONE submit, hidden resident, one readback. Updates h in place. false = refusal. loop_norm_at: virtual layer indices after which final_norm is applied (Looped Transformer mid-stack norm). Empty for standard models.