[cherrypick-2.3] modify infer gpu memory strategy (#41427), remove...

[cherrypick-2.3] modify infer gpu memory strategy (#41427), remove cudnn_deterministic=True (#41341)  (#41491)
Co-authored-by: NJingZhuangzhuang <75348594+JZZ-NOTE@users.noreply.github.com>

paddle/fluid/inference/api/analysis_predictor.cc

@@ -1058,16 +1058,8 @@ std::unique_ptr<PaddlePredictor> CreatePaddlePredictor<
                            std::to_string(fraction_of_gpu_memory);
         VLOG(3) << "set flag: " << flag;
         gflags.push_back(flag);
-        gflags.push_back("--cudnn_deterministic=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()) {

paddle/fluid/platform/flags.cc

@@ -364,11 +364,7 @@ PADDLE_DEFINE_EXPORTED_double(
  * Example:
  * Note: For selecting allocator policy of PaddlePaddle.
  */
-#ifdef PADDLE_ON_INFERENCE
-static constexpr char kDefaultAllocatorStrategy[] = "naive_best_fit";
-#else
 static constexpr char kDefaultAllocatorStrategy[] = "auto_growth";
-#endif
 PADDLE_DEFINE_EXPORTED_string(
     allocator_strategy, kDefaultAllocatorStrategy,
     "The allocation strategy, enum in [naive_best_fit, auto_growth]. "