[AutoParallel] Fix pipeline visualization tool (cherry-pick from f086447) (#71695)

* fix profile tool

Author: AndSonder

paddle/fluid/framework/new_executor/pir_interpreter.cc

@@ -1939,7 +1939,7 @@ void PirInterpreter::RunInstructionBase(InstructionBase* instr_node) {
       ::pir::Operation* op = instr_node->Operation();
       if (!calculate_stream_timer_->IsStarted() && op_name != "pd_op.feed" &&
           !op->HasAttribute("ring_id") && op_name != "pd_op.shadow_feed" &&
-          op_name != "pd_op.full_int_array") {
+          op_name != "pd_op.full" && op_name != "pd_op.full_int_array") {
         VLOG(3) << "Start calculated stream timer from op: " << op_name;
         calculate_stream_timer_->Start();
       }

python/paddle/distributed/auto_parallel/static/profiler_helper_static.py

@@ -150,6 +150,7 @@ def process_one_machine_log(log_dir, multi_machine_idx=-1):
                     step_infos.append([float("inf"), float("-inf")])
                 step_infos[i][0] = min(step_infos[i][0], info[0])
                 step_infos[i][1] = max(step_infos[i][1], info[1])
+        return start_step
 
     if args.multi_machine:
         multi_machine_dirs = os.listdir(args.log_dir)
@@ -162,9 +163,9 @@ def process_one_machine_log(log_dir, multi_machine_idx=-1):
         machine_num = len(multi_machine_dirs)
         for i, d in enumerate(multi_machine_dirs):
             _logger.info(f"Process machine {i}")
-            process_one_machine_log(d, i)
+            start_step = max(process_one_machine_log(d, i), start_step)
     else:
-        process_one_machine_log(args.log_dir)
+        start_step = process_one_machine_log(args.log_dir)
 
     for i, info in enumerate(step_infos):
         start_time = info[0]

python/paddle/profiler/utils.py

@@ -15,6 +15,7 @@
 from __future__ import annotations
 
 import functools
+import logging
 import sys
 from contextlib import ContextDecorator, contextmanager
 from typing import TYPE_CHECKING
@@ -271,6 +272,9 @@ def job_schedule_profiler_range(iter_id, start, end, exit_after_prof=True):
 def switch_job_schedule_profiler(
     model, iter_id, start, end, exit_after_prof=True
 ):
+    logging.info(
+        f"Schedule Profiler start at step {start} and end at step {end}"
+    )
     with job_schedule_profiler_range(
         iter_id, start, end, exit_after_prof
     ) as status: