Struct MultiViewTransformerBlock 

pub struct MultiViewTransformerBlock { /* private fields */ }

A transformer block with multi-view cross-attention support.

Each block contains:

1. Self-attention (within each view)
2. Cross-view attention (across all N views)
3. Text/prompt cross-attention
4. IP cross-attention (reference image CLIP embedding)
5. Feed-forward network

Implementations

impl MultiViewTransformerBlock

pub fn new( vs: VarBuilder<'_>, dim: usize, n_heads: usize, d_head: usize, context_dim: usize, ip_dim: usize, num_views: usize, ) -> Result<Self>

Create a new multi-view transformer block with standard attention.

pub fn new_with_flash( vs: VarBuilder<'_>, dim: usize, n_heads: usize, d_head: usize, context_dim: usize, ip_dim: usize, num_views: usize, use_flash_attention: bool, flash_block_size: usize, ) -> Result<Self>

Create a new multi-view transformer block with optional flash attention.

Arguments

• vs - Variable builder for weight initialization
• dim - Hidden dimension
• n_heads - Number of attention heads
• d_head - Dimension per head
• context_dim - Text cross-attention context dimension
• ip_dim - IP-adapter context dimension
• num_views - Number of views for cross-view attention
• use_flash_attention - Whether to use flash attention
• flash_block_size - Block size for flash attention tiling

pub fn forward( &self, xs: &Tensor, context: Option<&Tensor>, ip_tokens: Option<&Tensor>, ) -> Result<Tensor>

Forward pass.

• xs: (B*num_views, seq_len, dim) — spatial tokens for all views (batched)
• context: (B*num_views, ctx_len, context_dim) — text encoder hidden states
• ip_tokens: (B*num_views, ip_len, ip_dim) — CLIP image embedding tokens

Trait Implementations

impl Debug for MultiViewTransformerBlock

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

Formats the value using the given formatter.