diff --git a/deep_gemm/include/deep_gemm/impls/sm100_tf32_hc_prenorm_gemm.cuh b/deep_gemm/include/deep_gemm/impls/sm100_tf32_hc_prenorm_gemm.cuh
index aaf7fd9aea..47b68e978a 100644
--- a/deep_gemm/include/deep_gemm/impls/sm100_tf32_hc_prenorm_gemm.cuh
+++ b/deep_gemm/include/deep_gemm/impls/sm100_tf32_hc_prenorm_gemm.cuh
@@ -2,6 +2,8 @@
 #pragma clang diagnostic push
 #pragma clang diagnostic ignored "-Wunknown-attributes"
 
+#include <cuda_device_runtime_api.h>
+
 #include <cutlass/arch/barrier.h>
 
 #include <deep_gemm/common/cute_tie.cuh>
diff --git a/deep_gemm/include/deep_gemm/impls/sm90_tf32_hc_prenorm_gemm.cuh b/deep_gemm/include/deep_gemm/impls/sm90_tf32_hc_prenorm_gemm.cuh
index 93b1410010..f0e981387a 100644
--- a/deep_gemm/include/deep_gemm/impls/sm90_tf32_hc_prenorm_gemm.cuh
+++ b/deep_gemm/include/deep_gemm/impls/sm90_tf32_hc_prenorm_gemm.cuh
@@ -2,6 +2,8 @@
 #pragma clang diagnostic push
 #pragma clang diagnostic ignored "-Wunknown-attributes"
 
+#include <cuda_device_runtime_api.h>
+
 #include <cutlass/arch/barrier.h>
 #include <cutlass/arch/reg_reconfig.h>