make a copy of the index on CPU

source/sample/transformer/T2TSearch.cpp

@@ -252,9 +252,14 @@ void T2TSearch::Generate(T2TStateBundle * beam)
     InitTensor(&score, &scoreTopK);
     CopyValues(scoreTopK, score);
 
+    /*  CPU data (TODO: remove GPU->CPU data copy!!!) */
+    XTensor indexCPU;
+    InitTensor(&indexCPU, index.order, index.dimSize, index.dataType, index.denseRatio, -1);
+    CopyValues(index, indexCPU);
+
     /* sequence probability of top-k candidates */
     InitTensor(&probPath, &scoreTopK);
-    _Gather(&beam->prob, &probPath, probPath.order - 1, (int*)index.data, index.unitNum);
+    _Gather(&beam->prob, &probPath, probPath.order - 1, (int*)indexCPU.data, indexCPU.unitNum);
 }
 
 /*