[DCU] fix compile

transformer_engine/common/fused_router/fused_moe_aux_loss.cu

@@ -5,8 +5,8 @@
  ************************************************************************/
 
 #include <assert.h>
-#include <cooperative_groups.h>
 #include <cuda_runtime.h>
+#include <cooperative_groups.h>
 #include <transformer_engine/fused_router.h>
 
 #include "../common.h"

transformer_engine/common/fused_router/utils.h

@@ -9,6 +9,10 @@
 
 #include "transformer_engine/transformer_engine.h"
 
+#ifdef __HIP_PLATFORM_AMD__
+#define __syncwarp __syncthreads
+#endif
+
 namespace transformer_engine {
 
 constexpr size_t kThreadsPerWarp = 32;