add check that K_q must be a multiple of 2

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

@@ -1152,6 +1152,13 @@ struct DeviceBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V1
             return false;
         }
 
+        // saving dQ data with atomic_add instruction, so KzRaw must be a multiple of 2
+        if(KzRaw % 2 != 0)
+        {
+            std::cout << "K_q must be a multiple of 2" << std::endl;
+            return false;
+        }
+
         // Check vector load/store requirement
         const auto a_stride_lowest =
             ABlockTransferSrcVectorDim == 2 ? arg.a_mz_kz_strides_[1] : arg.a_mz_kz_strides_[0];

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

@@ -1189,6 +1189,13 @@ struct DeviceBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V2
             return false;
         }
 
+        // saving dQ data with atomic_add instruction, so KzRaw must be a multiple of 2
+        if(KzRaw % 2 != 0)
+        {
+            std::cout << "K_q must be a multiple of 2" << std::endl;
+            return false;
+        }
+
         // Check vector load/store requirement
         const auto a_stride_lowest =
             ABlockTransferSrcVectorDim == 2 ? arg.a_mz_kz_strides_[1] : arg.a_mz_kz_strides_[0];

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

@@ -1181,6 +1181,13 @@ struct DeviceGroupedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V1
                 return false;
             }
 
+            // saving dQ data with atomic_add instruction, so KzRaw must be a multiple of 2
+            if(KzRaw % 2 != 0)
+            {
+                std::cout << "K_q must be a multiple of 2" << std::endl;
+                return false;
+            }
+
             // Check vector load/store requirement
             const auto a_stride_lowest = ABlockTransferSrcVectorDim == 2
                                              ? device_arg.a_mz_kz_strides_[1]

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

@@ -1252,6 +1252,13 @@ struct DeviceGroupedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V2
                 return false;
             }
 
+            // saving dQ data with atomic_add instruction, so KzRaw must be a multiple of 2
+            if(KzRaw % 2 != 0)
+            {
+                std::cout << "K_q must be a multiple of 2" << std::endl;
+                return false;
+            }
+
             // Check vector load/store requirement
             const auto a_stride_lowest = ABlockTransferSrcVectorDim == 2
                                              ? device_arg.a_mz_kz_strides_[1]