polish allocator strategy doc, test=develop, test=document_fix (#22095)

paddle/fluid/platform/flags.cc

@@ -312,11 +312,19 @@ 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. "
-              "Enum in [naive_best_fit, auto_growth].");
+DEFINE_string(
+    allocator_strategy, kDefaultAllocatorStrategy,
+    "The allocation strategy, enum in [naive_best_fit, auto_growth]. "
+    "naive_best_fit means the original pre-allocated allocator of Paddle. "
+    "auto_growth means the auto-growth allocator. "
+    "These two strategies differ in GPU memory allocation. "
+    "naive_best_fit strategy would occupy almost all GPU memory by default, "
+    "which prevents users from starting several Paddle jobs on the same GPU "
+    "card but leads to less memory fragmentation (i.e., maximum batch "
+    "size of models may be larger). auto_growth strategy would allocate "
+    "GPU memory on demand, which allows users to start several Paddle jobs "
+    "on the same GPU card but may lead to more memory fragmentation "
+    "(i.e., maximum batch size of models may be smaller).");
 
 /**
  * Memory related FLAG