pub struct AttentionOptions {
pub causal: bool,
pub causal_offset: u32,
pub window: u32,
pub dropout_prob: f32,
pub seed: u64,
}Expand description
Options for ComputeEngine::flash_attention_v2_with and
ComputeEngine::flash_attention_v2_backward_with.
Query i sits at absolute position i + causal_offset; key j is
visible when j <= i + causal_offset (if causal) and
|i + causal_offset - j| < window (if window > 0).
Fields§
§causal: boolHide keys after each query’s absolute position.
causal_offset: u32Absolute position of the first query, e.g. the number of tokens already in a KV cache when decoding.
window: u32Sliding-window size in tokens; 0 means unlimited.
dropout_prob: f32Attention dropout probability; 0.0 disables dropout.
seed: u64Seed for the dropout pattern (only the low 32 bits are used).
Trait Implementations§
Source§impl Clone for AttentionOptions
impl Clone for AttentionOptions
impl Copy for AttentionOptions
Source§impl Debug for AttentionOptions
impl Debug for AttentionOptions
Auto Trait Implementations§
impl Freeze for AttentionOptions
impl RefUnwindSafe for AttentionOptions
impl Send for AttentionOptions
impl Sync for AttentionOptions
impl Unpin for AttentionOptions
impl UnsafeUnpin for AttentionOptions
impl UnwindSafe for AttentionOptions
Blanket Implementations§
Source§impl<T> BorrowMut<T> for Twhere
T: ?Sized,
impl<T> BorrowMut<T> for Twhere
T: ?Sized,
Source§fn borrow_mut(&mut self) -> &mut T
fn borrow_mut(&mut self) -> &mut T
Mutably borrows from an owned value. Read more