diff --git a/mlx/backend/cuda/scaled_dot_product_attention.cu b/mlx/backend/cuda/scaled_dot_product_attention.cu
index 9af6160d3..903b7c4b6 100644
--- a/mlx/backend/cuda/scaled_dot_product_attention.cu
+++ b/mlx/backend/cuda/scaled_dot_product_attention.cu
@@ -1068,8 +1068,7 @@ void ScaledDotProductAttention::eval_gpu(
           flags);
     }
 
-    // return sdpa_vector_fallback(s, encoder, q, k, v, scale_, o, do_causal_);
-    return sdpa_cudnn(s, encoder, q, k, v, scale_, o, do_causal_);
+    return sdpa_vector_fallback(s, encoder, q, k, v, scale_, o, do_causal_);
   }
 
   // Full attention mode