diff --git a/onnxruntime/test/contrib_ops/attention_op_test_helper.cc b/onnxruntime/test/contrib_ops/attention_op_test_helper.cc
index bac91aa544..44a9808119 100644
--- a/onnxruntime/test/contrib_ops/attention_op_test_helper.cc
+++ b/onnxruntime/test/contrib_ops/attention_op_test_helper.cc
@@ -7,7 +7,8 @@
 namespace onnxruntime {
 namespace test {
 
-#if !defined(_MSC_VER) || defined(USE_DML)
+// Disable some tests in Windows since prefast build might crash with large test data.
+#if !defined(_MSC_VER)
 void GetWeight_64_3_64(std::vector<float>& weight_data) {
   weight_data = {
       -0.004707f, -0.006775f, 0.0009236f, 0.003067f, -0.00806f, 0.00779f, 0.0004425f, 0.00846f, 0.00048f,
diff --git a/onnxruntime/test/contrib_ops/attention_op_test_helper.h b/onnxruntime/test/contrib_ops/attention_op_test_helper.h
index 0e2241e23e..aa2a371cc9 100644
--- a/onnxruntime/test/contrib_ops/attention_op_test_helper.h
+++ b/onnxruntime/test/contrib_ops/attention_op_test_helper.h
@@ -43,7 +43,7 @@ struct AttentionTestData {
 };
 
 // Disable some tests in Windows since prefast build might crash with large test data.
-#if !defined(_MSC_VER) || defined(USE_DML)
+#if !defined(_MSC_VER)
 // Return packed weights and bias for input projection.
 void GetAttentionWeight(std::vector<float>& weight_data, int elements = 64 * 3 * 64, int offset = 0, int step = 1);
 void GetAttentionBias(std::vector<float>& bias_data, int elements = 3 * 64, int offset = 0, int step = 1);
diff --git a/onnxruntime/test/contrib_ops/multihead_attention_op_test.cc b/onnxruntime/test/contrib_ops/multihead_attention_op_test.cc
index 5cd42e815f..fc7eafd714 100644
--- a/onnxruntime/test/contrib_ops/multihead_attention_op_test.cc
+++ b/onnxruntime/test/contrib_ops/multihead_attention_op_test.cc
@@ -452,7 +452,8 @@ static void RunMultiHeadAttentionTests(AttentionTestData& data, bool disable_cpu
   }
 }
 
-#if !defined(_MSC_VER) || defined(USE_DML)
+// Disable some tests in Windows since prefast build might crash with large test data.
+#if !defined(_MSC_VER)
 // Test fused cross attention kernel
 // It requires head_size > 32 and head_size <= 64 for T4 GPU; hidden_size == v_hidden_size.
 TEST(MultiHeadAttentionTest, CrossAttention_Batch2_HeadSize40) {