Enum AttentionMode 

#[repr(u8)]
pub enum AttentionMode {
    Causal = 0,
    Bidirectional = 1,
    BlockCausal = 2,
    SpKv = 3,
    SpKvQuant = 4,
    DashAttn = 5,
}

Attention mode for forward passes.

• Causal: Standard autoregressive — only attend to positions ≤ current (default).
• Bidirectional: Attend to ALL positions — used for dLLM masked prediction (Plan 066).
• BlockCausal: Bidirectional within current block, causal across blocks — D2F student.
• SpKv: SP-KV self-pruned key-value attention (Plan 070).
• SpKvQuant: SP-KV + Quantized KV fusion (Plan 070 Phase 3, Task T12).

Variants

Causal = 0

Bidirectional = 1

Full bidirectional: all positions see all positions (teacher mode).

BlockCausal = 2

Block-causal: bidirectional within block, causal across blocks (student mode).

SpKv = 3

SP-KV self-pruned key-value attention (Plan 070). Learns which KV pairs to retain via utility prediction. Gate bias = log(u) during training, 0|-inf during inference.

SpKvQuant = 4

SP-KV + Quantized KV fusion (Plan 070 Phase 3, Task T12). Selective write (SP-KV utility gating) + lossy quantize (any QuantizedKVCache backend). Two-stage compression: only useful KV pairs kept, those compressed to 2-4 bits/coord.

DashAttn = 5

DashAttention: adaptive sparse hierarchical attention via α-entmax routing (Plan 106). Replaces fixed-budget top-k block selection with adaptive support selection. Learned chunk summaries via head_cls vectors.

Trait Implementations

impl Clone for AttentionMode

fn clone(&self) -> AttentionMode

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Copy for AttentionMode

impl Debug for AttentionMode

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for AttentionMode

fn default() -> AttentionMode

Returns the “default value” for a type.

impl Eq for AttentionMode

impl PartialEq for AttentionMode

fn eq(&self, other: &AttentionMode) -> bool

Equality operator ==.

fn ne(&self, other: &Rhs) -> bool

Inequality operator !=.

impl StructuralPartialEq for AttentionMode