embed_overlay_routed

metrale_model_layers::layers::ops::token_overlay

Function embed_overlay_routed 

Source 
pub fn embed_overlay_routed(
    gpu: &dyn GpuBackend,
    kernel: KernelHandle,
    ids: DevicePtr,
    seq_slot: DevicePtr,
    active: i32,
    slot_map_tab: DevicePtr,
    rows_tab: DevicePtr,
    n_tab: DevicePtr,
    out: DevicePtr,
    num_tokens: u32,
    h: u32,
    vocab: u32,
    stream: u64,
) -> Result<()>
Expand description

2026-09-25: In-place replacement of overridden vocab rows in out ([num_tokens, h] BF16), after the embed gather.

Per row r: s = seq_slot[r], or active when seq_slot is null; skip when s < 0, when slot_map_tab[s] is null, or when ids[r] >= vocab; slot = slot_map_tab[s][ids[r]]; skip when slot < 0 or slot >= n_tab[s]; otherwise copy rows_tab[s][slot] over out[r]. ids is [num_tokens] u32; the three tables hold one entry per adapter slot; vocab is the length of each slot map.