Implement attention masking in LabelAttentionClassifier cross-attention

[Copilot]

Addresses review feedback that LabelAttentionClassifier was ignoring the attention mask, allowing padding tokens to influence label embeddings and attention visualizations.

Changes

• Added attention_mask parameter to LabelAttentionClassifier.forward() and passed it from TextEmbedder._get_sentence_embedding()

• Masked cross-attention computation: Convert mask from (1=real, 0=pad) to boolean format, expand to (B, 1, 1, T) for broadcasting, and pass to F.scaled_dot_product_attention as attn_mask

• Masked attention matrix for explainability: Apply masked_fill(-inf) before softmax to zero out padding positions in attention weights

# Prepare mask for scaled_dot_product_attention
attn_mask = None
if attention_mask is not None:
    attn_mask = (attention_mask == 0)  # Convert: 0 (pad) -> True (mask out)
    attn_mask = attn_mask.unsqueeze(1).unsqueeze(2)  # (B, T) -> (B, 1, 1, T)

y = F.scaled_dot_product_attention(q, k, v, attn_mask=attn_mask, ...)

# For attention matrix computation
if attention_mask is not None:
    attention_scores = attention_scores.masked_fill(attn_mask, float('-inf'))
attention_matrix = torch.softmax(attention_scores, dim=-1)

---

💡 You can make Copilot smarter by setting up custom instructions, customizing its development environment and configuring Model Context Protocol (MCP) servers. Learn more Copilot coding agent tips in the docs.