Struct AttnFacts 

pub struct AttnFacts {
    pub attn_layer_idx: usize,
    pub kv_dtype: KvCacheDtype,
    pub num_q_heads: u32,
    pub num_kv_heads: u32,
    pub head_dim: u32,
    pub gated: bool,
    pub rope: RopeFacts,
    pub sliding_window: u32,
    pub softmax_scale: f32,
    pub paged_decode_plain_rows: u128,
}

2026-09-28: What a full-attention layer’s kernels need beyond its weights.

Fields

attn_layer_idx: usize

2026-09-28: This layer’s index among the attention layers: its KV-cache layer.

kv_dtype: KvCacheDtype

2026-09-28: KV-cache dtype of this layer.

num_q_heads: u32

2026-09-28: Query heads.

num_kv_heads: u32

2026-09-28: KV heads.

head_dim: u32

2026-09-28: Head dim.

gated: bool

2026-09-28: Q carries an output gate ([Q | Gate] after the split).

rope: RopeFacts

2026-09-28: Rotary embedding.

sliding_window: u32

2026-09-28: Sliding window in positions; 0 for none.

softmax_scale: f32

2026-09-28: Softmax scale (effective_attn_scale).

paged_decode_plain_rows: u128

2026-09-28: Bit r - 1 is set when the layer’s own decode routing at r rows picks the plain non-split paged kernel (paged_decode_k): no split-K pair, no GQA-packed kernel, no 512-wide head kernel.

Implementations

impl AttnFacts

pub fn paged_decode_plain(&self, rows: u64) -> bool

2026-09-28: The plain paged kernel serves rows rows (1..=128).

Trait Implementations

impl Clone for AttnFacts

fn clone(&self) -> AttnFacts

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for AttnFacts

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

Formats the value using the given formatter.

impl PartialEq for AttnFacts

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

Tests for self and other values to be equal, and is used by ==.

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

Tests for !=. The default implementation is almost always sufficient, and should not be overridden without very good reason.

impl Copy for AttnFacts

impl StructuralPartialEq for AttnFacts