Struct TwoWayAttentionBlock 

pub struct TwoWayAttentionBlock { /* private fields */ }

A two-way attention block: tokens attend to themselves and to the image, then the image attends back to the tokens — so both representations are updated in a single block.

Implementations

impl TwoWayAttentionBlock

pub fn new( embed_dim: usize, n_heads: usize, mlp_dim: usize, rng: &mut LcgRng, ) -> VisionResult<Self>

Construct a block.

Errors

Propagates MultiHeadAttention::new validation.

pub fn forward( &self, tokens: &[f32], image: &[f32], query_pe: &[f32], key_pe: &[f32], ) -> VisionResult<TwoWayBlockOutput>

Forward pass.

tokens/query_pe are [n_t · e]; image/key_pe are [n_i · e]. The positional encodings are added to queries and keys only (never to the values), exactly as in SAM.

Errors

• VisionError::DimensionMismatch on inconsistent lengths.
• Propagates attention errors.