diff --git a/onnxruntime/contrib_ops/cuda/bert/attention_impl.cu b/onnxruntime/contrib_ops/cuda/bert/attention_impl.cu
index 695cc2200fe7a..187f1bb37edc5 100644
--- a/onnxruntime/contrib_ops/cuda/bert/attention_impl.cu
+++ b/onnxruntime/contrib_ops/cuda/bert/attention_impl.cu
@@ -417,10 +417,10 @@ Status QkvToContext(
   const bool past_present_share_buffer = parameters.past_present_share_buffer;
   const float mask_filter_value = parameters.mask_filter_value;
   void* fused_runner = data.fused_runner;
-  bool use_memory_efficient_attention = data.use_memory_efficient_attention;
 
   // At most one fused kernel is enabled.
-  assert(int(use_memory_efficient_attention) + int(fused_runner != nullptr) + int(data.fused_cross_attention_kernel != nullptr) <= 1);
+  assert(int(data.use_memory_efficient_attention) + int(fused_runner != nullptr) +
+         int(data.fused_cross_attention_kernel != nullptr) <= 1);
 
   const int batches = batch_size * num_heads;
   const int size_per_batch_q = sequence_length * qk_head_size;
@@ -469,7 +469,7 @@ Status QkvToContext(
     assert(data.fused_cross_attention_kernel == nullptr);
     assert(!use_fused_kernel);
     assert(data.gemm_buffer != nullptr);
-    assert(!use_memory_efficient_attention);
+    assert(!data.use_memory_efficient_attention);
 
     if (data.present != data.past) {
       // For easy testing. Production should better avoid this path.
@@ -564,7 +564,7 @@ Status QkvToContext(
   }
 
 #if USE_FLASH_ATTENTION
-  if (use_memory_efficient_attention) {
+  if (data.use_memory_efficient_attention) {
     // We only enable fused cross attention when there is no key padding mask.
     // Otherwise, key have effective batch size 2 * batch_size, which is different from batch_size of query.
     assert(data.mask_index == nullptr);