Struct AttentionConfig 

pub struct AttentionConfig {
    pub batch_size: usize,
    pub num_heads: usize,
    pub query_len: usize,
    pub kv_len: usize,
    pub head_dim: usize,
    pub causal: bool,
    pub scale: f32,
    pub precision: KernelPrecision,
    pub block_q: usize,
    pub block_kv: usize,
}

Configuration for standard attention.

Fields

batch_size: usize

Batch size.

num_heads: usize

Number of attention heads.

query_len: usize

Query sequence length.

kv_len: usize

Key/Value sequence length.

head_dim: usize

Head dimension.

causal: bool

Whether to apply causal masking.

scale: f32

Softmax scale factor (usually 1/sqrt(head_dim)).

precision: KernelPrecision

Computation precision.

block_q: usize

Block size for query tiling.

block_kv: usize

Block size for KV tiling.

Implementations

impl AttentionConfig

pub fn new( batch_size: usize, num_heads: usize, query_len: usize, kv_len: usize, head_dim: usize, ) -> Self

Create config for the given dimensions.

pub fn with_causal(self, causal: bool) -> Self

Set causal masking.

pub fn with_precision(self, precision: KernelPrecision) -> Self

Set precision mode.

pub fn with_block_sizes(self, block_q: usize, block_kv: usize) -> Self

Set block sizes for tiling.

pub fn validate(&self) -> ConfigResult<()>

Validate configuration values.

Trait Implementations

impl Clone for AttentionConfig

fn clone(&self) -> AttentionConfig

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for AttentionConfig

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

Formats the value using the given formatter.

impl Default for AttentionConfig

fn default() -> Self

Returns the “default value” for a type.