Struct DecodeLayerInputs 

pub struct DecodeLayerInputs<'a> {
    pub x: &'a Tensor,
    pub k_cache: &'a Tensor,
    pub v_cache: &'a Tensor,
    pub cached_len_sym: SymId,
    pub attended_len_sym: SymId,
    pub offset: Option<&'a Tensor>,
    pub rope_cos: &'a Tensor,
    pub rope_sin: &'a Tensor,
    pub mask: &'a Tensor,
    pub attn_window: Option<usize>,
}

The per-layer arguments of DecodeBackbone::decode_apply_layer, bundled so the signature stays readable as families are added.

Fields

x: &'a Tensor

k_cache: &'a Tensor

v_cache: &'a Tensor

cached_len_sym: SymId

attended_len_sym: SymId

offset: Option<&'a Tensor>

Some on the device-offset (Op::WriteSliceDoff) path, None on the backend-generic SymEnv path. The layer must honour both — see the divergence note at the top of this module.

rope_cos: &'a Tensor

rope_sin: &'a Tensor

mask: &'a Tensor

This layer’s mask variant, already sliced out of the stacked Const.

attn_window: Option<usize>

This layer’s window width, or None if it attends densely — the same plan entry mask was built from (MaskPlan::window_for_layer).

Carried so a layer offering the CUDA flash-decode arm can state the truth about its own key range instead of asserting None. The arm’s flash_decoding kernel does not implement local attention, so a windowed layer’s offer is declined — which is correct and is the point. Asserting None there would make the arm attend the whole prefix and silently drop the window (GAP-194).