diff --git a/ex_engine/xllm_kernels/cuda/bindings/hgemm_bind.cpp b/ex_engine/xllm_kernels/cuda/bindings/hgemm_bind.cpp
index af94c952..f1a6028e 100644
--- a/ex_engine/xllm_kernels/cuda/bindings/hgemm_bind.cpp
+++ b/ex_engine/xllm_kernels/cuda/bindings/hgemm_bind.cpp
@@ -5,6 +5,7 @@
 //   moe_expert_gemm(input, weights, expert_counts) → output
 
 #include <torch/extension.h>
+#include <cuda_runtime.h>
 #include <cuda_fp16.h>
 #include <vector>