add #include <numeric>

include/ck/tensor_operation/gpu/device/impl/device_batched_mha_bwd_xdl_cshuffle_kloop_v1.hpp

@@ -5,6 +5,7 @@
 
 #include <iostream>
 #include <sstream>
+#include <numeric>
 
 #include "ck/utility/common_header.hpp"
 #include "ck/utility/philox_rand.hpp"

include/ck/tensor_operation/gpu/device/impl/device_batched_mha_bwd_xdl_cshuffle_kloop_v2.hpp

@@ -5,6 +5,7 @@
 
 #include <iostream>
 #include <sstream>
+#include <numeric>
 
 #include "ck/utility/common_header.hpp"
 #include "ck/utility/philox_rand.hpp"

include/ck/tensor_operation/gpu/device/impl/device_batched_mha_bwd_xdl_cshuffle_qloop_light_v1.hpp

@@ -5,6 +5,7 @@
 
 #include <iostream>
 #include <sstream>
+#include <numeric>
 
 #include "ck/utility/common_header.hpp"
 #include "ck/utility/philox_rand.hpp"

include/ck/tensor_operation/gpu/device/impl/device_batched_mha_bwd_xdl_cshuffle_qloop_light_v2.hpp

@@ -5,6 +5,7 @@
 
 #include <iostream>
 #include <sstream>
+#include <numeric>
 
 #include "ck/utility/common_header.hpp"
 #include "ck/utility/philox_rand.hpp"

include/ck/tensor_operation/gpu/device/impl/device_batched_mha_bwd_xdl_cshuffle_qloop_v1.hpp

@@ -5,6 +5,7 @@
 
 #include <iostream>
 #include <sstream>
+#include <numeric>
 
 #include "ck/utility/common_header.hpp"
 #include "ck/utility/philox_rand.hpp"

include/ck/tensor_operation/gpu/device/impl/device_batched_mha_bwd_xdl_cshuffle_qloop_v2.hpp

@@ -5,6 +5,7 @@
 
 #include <iostream>
 #include <sstream>
+#include <numeric>
 
 #include "ck/utility/common_header.hpp"
 #include "ck/utility/philox_rand.hpp"

include/ck/tensor_operation/gpu/device/impl/device_grouped_mha_bwd_xdl_cshuffle_kloop_v1.hpp

@@ -5,6 +5,7 @@
 
 #include <iostream>
 #include <sstream>
+#include <numeric>
 
 #include "ck/utility/common_header.hpp"
 #include "ck/utility/philox_rand.hpp"

include/ck/tensor_operation/gpu/device/impl/device_grouped_mha_bwd_xdl_cshuffle_kloop_v2.hpp

@@ -5,6 +5,7 @@
 
 #include <iostream>
 #include <sstream>
+#include <numeric>
 
 #include "ck/utility/common_header.hpp"
 #include "ck/utility/philox_rand.hpp"

include/ck/tensor_operation/gpu/device/impl/device_grouped_mha_bwd_xdl_cshuffle_qloop_light_v1.hpp

@@ -5,6 +5,7 @@
 
 #include <iostream>
 #include <sstream>
+#include <numeric>
 
 #include "ck/utility/common_header.hpp"
 #include "ck/utility/philox_rand.hpp"

include/ck/tensor_operation/gpu/device/impl/device_grouped_mha_bwd_xdl_cshuffle_qloop_light_v2.hpp

@@ -5,6 +5,7 @@
 
 #include <iostream>
 #include <sstream>
+#include <numeric>
 
 #include "ck/utility/common_header.hpp"
 #include "ck/utility/philox_rand.hpp"

include/ck/tensor_operation/gpu/device/impl/device_grouped_mha_bwd_xdl_cshuffle_qloop_v1.hpp

@@ -5,6 +5,7 @@
 
 #include <iostream>
 #include <sstream>
+#include <numeric>
 
 #include "ck/utility/common_header.hpp"
 #include "ck/utility/philox_rand.hpp"

include/ck/tensor_operation/gpu/device/impl/device_grouped_mha_bwd_xdl_cshuffle_qloop_v2.hpp

@@ -5,6 +5,7 @@
 
 #include <iostream>
 #include <sstream>
+#include <numeric>
 
 #include "ck/utility/common_header.hpp"
 #include "ck/utility/philox_rand.hpp"