Enum FlashAttnPrefillLayout 

pub enum FlashAttnPrefillLayout {
    HeadMajor,
    SeqMajor,
}

Layout selector for dispatch_flash_attn_prefill_bf16_d64.

The kernel reads from raw device pointers via integer strides, so any element layout that keeps head_dim (D) as the contiguous innermost axis is valid input. The two layouts named here both satisfy that constraint and cover every BERT/embedding caller in hf2q today:

• HeadMajor — [B, H, L, D], the same layout the D=256/D=512 dispatchers assume. Stride math: seq = D, head = L * D, batch = H * L * D.

• SeqMajor — [B, L, H, D], the natural output of BERT linear projections (hidden = H * D row-major). Stride math: seq = H * D, head = D, batch = L * H * D. Choosing this layout avoids three host-side transpose dispatches per layer (Q + K + V) plus one for the output, which is the entire point of the D=64 dispatcher’s existence — the BERT family wins on dispatch count, not raw FA perf.

Variants

HeadMajor

[B, H, L, D] — same as the D=256/D=512 dispatchers.

SeqMajor

[B, L, H, D] — natural BERT/embedding-model layout.

Trait Implementations

impl Clone for FlashAttnPrefillLayout

fn clone(&self) -> FlashAttnPrefillLayout

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for FlashAttnPrefillLayout

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

Formats the value using the given formatter.

impl PartialEq for FlashAttnPrefillLayout

fn eq(&self, other: &FlashAttnPrefillLayout) -> 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 FlashAttnPrefillLayout

impl Eq for FlashAttnPrefillLayout

impl StructuralPartialEq for FlashAttnPrefillLayout