Set waiterType of kGpuSync to kCUDA (#50241)

paddle/fluid/framework/new_executor/interpreter/stream_analyzer.cc

@@ -431,8 +431,7 @@ void StreamAnalyzer::ShrinkEventInfo(
 
 platform::DeviceType StreamAnalyzer::GetWaiterType(
     const Instruction& instr) const {
-  if (instr.KernelType() == OpFuncType::kCpuSync ||
-      instr.KernelType() == OpFuncType::kGpuSync) {
+  if (instr.KernelType() == OpFuncType::kCpuSync) {
     return platform::kCPU;
   } else {
     if (platform::is_xpu_place(place_)) {

paddle/fluid/framework/new_executor/interpretercore.cc

@@ -858,8 +858,6 @@ void InterpreterCore::RunOperator(const Instruction& instr_node) {
                                        : var_scope_.GetMutableScope();
   VLOG(4) << "Start run " << place << " " << op->DebugStringEx(local_scope);
 
-  SetDeviceId(place);
-
 #ifdef PADDLE_WITH_ASCEND_CL
   if (platform::is_npu_place(place)) {
     // NOTE(wangxi): nan/inf cannot be detected on NPU by checking the
@@ -989,6 +987,8 @@ void InterpreterCore::RunInstruction(const Instruction& instr_node) {
   platform::RecordEvent instruction_event(
       op->Type(), platform::TracerEventType::Operator, 1);
 
+  SetDeviceId(instr_node.DeviceContext().GetPlace());
+
   try {
     instr_node.WaitEvent(place_);