better code of multi-gpu runs

source/sample/transformer/T2TEncoder.cpp

@@ -128,8 +128,6 @@ XTensor AttEncoder::Make(XTensor &input, XTensor &mask, bool skipInputRes)
 
             /* layer normalization */
             x = attLayerNorms[i].Make(res);
-
-            llnum = -1;
         }
 
         /* fnn */

source/tensor/XTensor.cpp

@@ -552,11 +552,17 @@ void XTensor::SetZeroAll(XStream * stream)
         if(devID >= 0){
 #ifdef USE_CUDA
             int size = sizeof(int) + (sizeof(int)+sizeof(DTYPE)) * unitNumNonZero;
+            
+            int devIDBackup = 0;
+            cudaGetDevice(&devIDBackup);
+            cudaSetDevice(devID);
 
             if(stream == NULL)
                 cudaMemset(data, 0, size);
             else
                 cudaMemsetAsync(data, 0, size, stream->stream);
+            
+            cudaSetDevice(devIDBackup);
 #endif
         }
         else
@@ -567,10 +573,16 @@ void XTensor::SetZeroAll(XStream * stream)
     else{
         if(devID >= 0){
 #ifdef USE_CUDA
+            int devIDBackup = 0;
+            cudaGetDevice(&devIDBackup);
+            cudaSetDevice(devID);
+            
             if(stream == NULL)
                 cudaMemset(data, 0, unitNum * unitSize);
             else
                 cudaMemsetAsync(data, 0, unitNum * unitSize, stream->stream);
+            
+            cudaSetDevice(devIDBackup);
 #endif
         }
         else