Module layer 

Composable transformer layer traits (SDD).

Decouples the generic model loop (embed -> layers -> norm -> lm_head) from layer-specific logic (attention vs SSM, MoE vs dense FFN). Adding a new architecture only requires implementing TransformerLayer for each layer type, not duplicating the model loop.

Structs

AttnLayerState

Attention-layer per-sequence state. KV lives in PagedKvCache; the only resident piece is the QSA indexer carry on the 12 qwen4_exp full-attention layers (Avarok #753 item B).

AttnMetadataDev

Pre-uploaded attention metadata device pointers.

BatchedAttnMetadata

Q12 batched-prefill device-side metadata.

EmptyLayerState

Empty state for layers that store all persistent state externally (e.g., attention layers where KV is in PagedKvCache).

ForwardContext

Shared context for a single forward pass step.

GdnPrefillBuffers

Device pointers to full-sequence GDN input/output buffers.

MidchunkCapture

Per-pass descriptor for mid-chunk SSM tail capture. Points at the reserved Marconi snapshot slot’s per-SSM-layer destination buffers (already offset to the slot) plus the split point in local (chunk) token coordinates.

SsmLayerState

SSM layer state: recurrent hidden state + conv1d sliding window.

Enums

MoeLoraRoute

Feature-1 MoE-LoRA fold decision for a single forward pass.

Constants

VERIFY_WY_LAYER_STRIDE_BYTES

Bytes between consecutive GDN layers’ table slices.

VERIFY_WY_TABLES_PER_LAYER

Tables per GDN layer: h_state + Hi0..Hi2 (K=4 verify → 3 intermediates).

VERIFY_WY_TABLE_SEQS

Batched-verify WY pointer-table layout (SSOT — writer: TransformerModel::upload_verify_wy_tables, reader: the qwen3_ssm GdnStates::Multi batched conv+WY arm).

VERIFY_WY_TABLE_STRIDE_BYTES

Bytes between consecutive tables within a layer slice.

Traits

LayerState

Per-layer persistent state tracked across decode steps.

TransformerLayer