export start event

dlrover/python/master/main.py

@@ -23,6 +23,7 @@
 from dlrover.python.master.args import parse_master_args
 from dlrover.python.scheduler.factory import new_job_args
 from dlrover.python.scheduler.job import JobArgs
+from dlrover.python.training_event import DLRoverMaster
 
 _dlrover_context = Context.singleton_instance()
 
@@ -62,6 +63,11 @@ def run(args):
 
 def main():
     args = parse_master_args()
+
+    # export event for dlrover master
+    master = DLRoverMaster()
+    master.start(pid=vars(args))
+
     exit_code = run(args)
     return exit_code
 

dlrover/python/training_event/predefined/dlrover.py

@@ -2,6 +2,7 @@
 from typing import Optional, List
 
 from .common import CommonPredefined
+from dlrover.python.common.singleton import Singleton
 
 
 class DLRoverCommonEventName(Enum):
@@ -133,7 +134,7 @@ def exit(self, reason: str, **kwargs):
         )
 
 
-class DLRoverMaster(DLRoverCommon):
+class DLRoverMaster(DLRoverCommon, Singleton):
     """DLRover Master events"""
 
     def __init__(self) -> None:
@@ -193,7 +194,7 @@ def fault_detect(self, reason: str, **kwargs):
         )
 
 
-class DLRoverAgent(DLRoverCommon):
+class DLRoverAgent(DLRoverCommon, Singleton):
     """DLRover Agent events"""
 
     def __init__(self) -> None:

dlrover/trainer/torch/elastic_run.py

@@ -120,6 +120,7 @@
     launch_agent,
 )
 from dlrover.trainer.torch.utils import version_less_than_230
+from dlrover.python.training_event import DLRoverAgent
 
 
 def parse_args(args):
@@ -446,6 +447,10 @@ def _check_to_use_dlrover_run(master_addr, max_nodes, timeout=120):
 
 
 def run(args):
+    # export event for dlrover agent
+    agent = DLRoverAgent()
+    agent.start(pid=vars(args))
+
     logger.info(f"DLRover agent started with: {cu.get_dlrover_version()}.")
     master_handler = None
     master_addr = os.getenv(NodeEnv.DLROVER_MASTER_ADDR, "")