diff --git a/onnxruntime/contrib_ops/cuda/bert/gqa_unfused_attention.cu b/onnxruntime/contrib_ops/cuda/bert/gqa_unfused_attention.cu
index f65a041a80630..8aac549aeba01 100644
--- a/onnxruntime/contrib_ops/cuda/bert/gqa_unfused_attention.cu
+++ b/onnxruntime/contrib_ops/cuda/bert/gqa_unfused_attention.cu
@@ -4,7 +4,7 @@
 // GQA-capable unfused CUDA attention kernel. See header for contract.
 
 #include <math_constants.h>
-#include <cub/cub.cuh>
+#include "core/providers/cuda/cu_inc/cub.cuh"
 #include <cuda_bf16.h>
 #include <cuda_fp16.h>
 #include "core/common/safeint.h"