Debug: Attention Scale

The function below is supposed to compute scaled dot-product attention, but it has a bug. Find and fix it.

Signature: def buggy_attention(Q, K, V)

• Q: (N, d_k) — query matrix
• K: (N, d_k) — key matrix
• V: (N, d_v) — value matrix
• Returns: (N, d_v)

Hint: The scaling factor is critical for numerical stability and proper gradient flow. Look carefully at how the attention scores are divided before the softmax.