Enum CausalAlignment 

pub enum CausalAlignment {
    BottomRight,
    TopLeft,
}

Anchor of the causal diagonal for AttentionModuleOptions::is_causal.

Query i may attend key j iff j <= i + offset, where offset is 0 for TopLeft and seq_k - seq_q for BottomRight.

Variants

BottomRight

The last query row sees every key: offset = seq_k - seq_q. This is the KV-cache convention (the queries are the newest seq_q positions), used by FlashAttention 2.1+ and by ONNX Attention when a past_key is supplied.

TopLeft

The first query row sees the first key: offset = 0 (tril on [seq_q, seq_k]). This is PyTorch’s scaled_dot_product_attention(is_causal=True) and ONNX Attention without a KV cache.

Implementations

impl CausalAlignment

pub fn offset(self, seq_q: usize, seq_k: usize) -> i64

The offset of the causal diagonal: query i may attend key j iff j <= i + offset.

Trait Implementations

impl Clone for CausalAlignment

fn clone(&self) -> Self

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Copy for CausalAlignment

impl Debug for CausalAlignment

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

Formats the value using the given formatter.

impl Default for CausalAlignment

fn default() -> Self

Returns the “default value” for a type.

impl<'de> Deserialize<'de> for CausalAlignment

fn deserialize<__D>(__deserializer: __D) -> Result<Self, __D::Error>

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl Eq for CausalAlignment

impl Hash for CausalAlignment

fn hash<__H: Hasher>(&self, state: &mut __H)

Feeds this value into the given Hasher.

fn hash_slice<H>(data: &[Self], state: &mut H)

where H: Hasher, Self: Sized,

Feeds a slice of this type into the given Hasher.

impl PartialEq for CausalAlignment

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

Equality operator ==.

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

Inequality operator !=.

impl Serialize for CausalAlignment

fn serialize<__S>(&self, __serializer: __S) -> Result<__S::Ok, __S::Error>

where __S: Serializer,

Serialize this value into the given Serde serializer.

impl StructuralPartialEq for CausalAlignment