add log while running New Executor, Old Executor and ParallelExecutor and change log level (#45814)

* optimize executor log

* delete log in new exe

* add log for old executor

* use LOG_FIRST_N(INFO, 1)

paddle/fluid/framework/executor.cc

@@ -183,6 +183,7 @@ void Executor::Run(const ProgramDesc& pdesc,
                    const std::vector<std::string>& skip_ref_cnt_vars,
                    bool force_disable_gc,
                    bool keep_kid_scopes) {
+  LOG_FIRST_N(INFO, 1) << "Old Executor is Running.";
   platform::RecordEvent record_run(
       "Executor::Run", platform::TracerEventType::UserDefined, 1);
   platform::RecordBlock b(block_id);

paddle/fluid/framework/new_executor/interpretercore.cc

@@ -193,6 +193,7 @@ paddle::framework::FetchList InterpreterCore::Run(
 #endif
 
   if (!is_build_) {
+    LOG_FIRST_N(INFO, 1) << "New Executor is Running.";
     paddle::framework::interpreter::build_variable_scope(
         block_, &var_scope_, create_local_scope_);
 

paddle/fluid/framework/new_executor/interpretercore_util.cc

@@ -418,7 +418,6 @@ void build_op_func_list(const platform::Place& place,
                                        : var_scope->GetMutableScope();
   std::vector<std::unique_ptr<OperatorBase>>
       ops_unique;  // its elements will be moved to vec_func_list
-  bool flag_log_is_printed = false;
   // Step 1: create all ops for current block.
   create_all_ops(block, &ops_unique);
 
@@ -443,6 +442,7 @@ void build_op_func_list(const platform::Place& place,
   }
   auto unused_var_map = get_unused_vars(block, ops);
 
+  bool flag_log_is_printed = false;
   for (size_t i = 0; i < ops.size(); ++i) {
     auto op = ops[i].get();
     const std::string& op_type = op->Type();
@@ -452,7 +452,7 @@ void build_op_func_list(const platform::Place& place,
     // Print new executor log if grad op is used.
     // It's only for test and will be removed later.
     if (!flag_log_is_printed && op_type.find("_grad") != std::string::npos) {
-      VLOG(0) << "Standalone Executor is Used.";
+      LOG_FIRST_N(INFO, 1) << "Standalone Executor is Used.";
       flag_log_is_printed = true;
     }
 

paddle/fluid/framework/parallel_executor.cc

@@ -983,6 +983,7 @@ void ParallelExecutor::BCastParamsToDevices(
 
 FetchUnmergedList ParallelExecutor::Run(
     const std::vector<std::string> &fetch_tensors) {
+  LOG_FIRST_N(INFO, 1) << "ParallelExecutor is Running (Run).";
   PreludeToRun(fetch_tensors);
   platform::RecordBlock b(0);
 
@@ -1000,6 +1001,7 @@ FetchUnmergedList ParallelExecutor::Run(
 
 FetchList ParallelExecutor::RunAndMerge(
     const std::vector<std::string> &fetch_tensors) {
+  LOG_FIRST_N(INFO, 1) << "ParallelExecutor is Running (RunAndMerge).";
   PreludeToRun(fetch_tensors);
   platform::RecordBlock b(0);
 

python/paddle/fluid/executor.py

@@ -771,9 +771,6 @@ class _ExecutorCache(object):
 
             inner_program = converted_program
             # print(f"Program after convert:\n {inner_program}", flush=True)
-            warnings.warn(
-                "FLAGS_USE_STANDALONE_EXECUTOR and FLAGS_CONVERT_GRAPH_TO_PROGRAM is set to 1. Graph will be converted to Program and executed using new executor."
-            )
         else:
             build_strategy = None
             from paddle.incubate.autograd import prim_enabled, prim2orig
@@ -789,9 +786,16 @@ class _ExecutorCache(object):
                                       fetch_var_name=fetch_var_name,
                                       use_fetch_v2=True)
 
-        # If there are multiple blocks in the program, subblock will not be executed with the new executor in temporary
-        if program.num_blocks > 1:
-            warnings.warn("There are more than 1 block in program.")
+        if os.environ.get('FLAGS_CONVERT_GRAPH_TO_PROGRAM', None) in [
+                1, '1', True, 'True', 'true'
+        ] and not program._is_start_up_program_:
+            if program.num_blocks > 1:
+                # If there are multiple blocks in the program, subblock will not be executed with the new executor in temporary
+                logging.warning("There are more than 1 block in program.")
+            elif program.num_blocks == 1:
+                logging.warning("There are 1 block in program.")
+            else:
+                logging.warning("There are no block in program.")
 
         # standalone executor will apply buffer_shared_inplace_pass and
         # inplace_addto_op_pass to program according to build_strategy
@@ -1667,10 +1671,6 @@ class Executor(object):
                 else:
                     tensor._copy_from(cpu_tensor, self.place)
 
-            warnings.warn(
-                "FLAGS_USE_STANDALONE_EXECUTOR is set to 1. New executor is used to execute Program."
-            )
-
             return new_exe.run(scope, list(feed.keys()), fetch_list,
                                return_numpy)