fix the jetson allocator strategy, test=develop (#32932)

paddle/fluid/inference/api/analysis_predictor.cc

@@ -650,20 +650,22 @@ std::unique_ptr<PaddlePredictor> CreatePaddlePredictor<
         gflags.push_back("--cudnn_deterministic=True");
       }
 
-      if (config.thread_local_stream_enabled()) {
-        gflags.push_back("--allocator_strategy=thread_local");
-        process_level_allocator_enabled = false;
-      } else {
-        process_level_allocator_enabled = true;
-      }
-
 // TODO(wilber): jetson tx2 may fail to run the model due to insufficient memory
 // under the native_best_fit strategy. Modify the default allocation strategy to
 // auto_growth. todo, find a more appropriate way to solve the problem.
 #ifdef WITH_NV_JETSON
       gflags.push_back("--allocator_strategy=auto_growth");
 #endif
 
+      // TODO(Shixiaowei02): Add a mandatory scheme to use the thread local
+      // allocator when multi-stream is enabled.
+      if (config.thread_local_stream_enabled()) {
+        gflags.push_back("--allocator_strategy=thread_local");
+        process_level_allocator_enabled = false;
+      } else {
+        process_level_allocator_enabled = true;
+      }
+
       if (framework::InitGflags(gflags)) {
         VLOG(3) << "The following gpu analysis configurations only take effect "
                    "for the first predictor: ";