diff --git a/paddle/fluid/memory/allocation/auto_growth_best_fit_allocator.cc b/paddle/fluid/memory/allocation/auto_growth_best_fit_allocator.cc
index f71a4b8e1a823da3ce6ec57fd39e21aa94a306a7..8cd7335cf4fa6dac77233b49c06053f4dc4190e6 100644
--- a/paddle/fluid/memory/allocation/auto_growth_best_fit_allocator.cc
+++ b/paddle/fluid/memory/allocation/auto_growth_best_fit_allocator.cc
@@ -27,7 +27,7 @@ DEFINE_bool(free_idle_chunk, false,
             "allocation request. If true, no allocation would be cached. This "
             "flag only works when FLAGS_allocator_strategy=auto_growth.");
 
-DEFINE_bool(free_when_no_cache_hit, true,
+DEFINE_bool(free_when_no_cache_hit, false,
             "Whether to free idle chunks when no cache hit. If true, idle "
             "chunk would be freed when no cache hit; if false, idle "
             "chunk would be freed when out of memory occurs. This flag "
diff --git a/paddle/fluid/platform/flags.cc b/paddle/fluid/platform/flags.cc
index 9b5e0c92fca470cad82191f41500120cecdf7f6f..f916567d99b1628443329744059194f0418997d5 100644
--- a/paddle/fluid/platform/flags.cc
+++ b/paddle/fluid/platform/flags.cc
@@ -307,7 +307,12 @@ DEFINE_double(memory_fraction_of_eager_deletion, 1.0,
  * Example:
  * Note: For selecting allocator policy of PaddlePaddle.
  */
-DEFINE_string(allocator_strategy, "naive_best_fit",
+#ifdef PADDLE_ON_INFERENCE
+static constexpr char kDefaultAllocatorStrategy[] = "naive_best_fit";
+#else
+static constexpr char kDefaultAllocatorStrategy[] = "auto_growth";
+#endif
+DEFINE_string(allocator_strategy, kDefaultAllocatorStrategy,
               "The allocation strategy. naive_best_fit means the original best "
               "fit allocator of Fluid. "
               "auto_growth means the experimental auto-growth allocator. "