pub fn load(
model: &Arc<CmfModel>,
cfg: &Dsv41Cfg,
n_layers: usize,
token_map: Vec<u32>,
) -> Result<(Dsv41Globals, Vec<Dsv41Layer>, Option<EngramHash>), String>Expand description
Build the stack from canonical V4.1 tensors. Frequency tables are kept separate because compressed latents use the 160k base plus YaRN while the raw sliding window uses the 10k base without YaRN.