Struct FlashAttentionConfig 

pub struct FlashAttentionConfig {
    pub block_size: usize,
    pub causal: bool,
    pub softmax_eps: f64,
}

Configuration for Flash Attention computation.

Fields

block_size: usize

Block size for tiled computation. Larger blocks use more memory but may be faster due to better cache utilization. Default: 64.

causal: bool

Whether to use causal masking (for autoregressive models). Default: false.

softmax_eps: f64

Epsilon for numerical stability in softmax. Default: 1e-6.

Implementations

impl FlashAttentionConfig

pub fn with_block_size(block_size: usize) -> Self

Create a new Flash Attention config with specified block size.

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

Enable causal masking for autoregressive attention.

Trait Implementations

impl Clone for FlashAttentionConfig

fn clone(&self) -> FlashAttentionConfig

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for FlashAttentionConfig

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

Formats the value using the given formatter.

impl Default for FlashAttentionConfig

fn default() -> Self

Returns the “default value” for a type.

impl Copy for FlashAttentionConfig