Use optimized attention in Dino2AttentionBlock.

comfy/image_encoders/dino2.py

@@ -53,8 +53,7 @@ class Dino2AttentionBlock(torch.nn.Module):
         if rope is not None and pos is not None:
             q = rope(q, pos)
             k = rope(k, pos)
-        out = F.scaled_dot_product_attention(q, k, v, attn_mask=mask)
-        out = out.transpose(1, 2).reshape(B, N, C)
+        out = optimized_attention(q, k, v, h, mask=mask, skip_reshape=True)
         return self.output(out)