Used rendesvous node as node_id in profiling event, enabling profiling multi-participants on one node.

Author: hexinw-nvidia

src/nvidia_resiliency_ext/fault_tolerance/_ft_rendezvous.py

@@ -1326,7 +1326,7 @@ def next_rendezvous(self) -> Union[RendezvousInfo, Tuple[Store, int, int]]:
         # Record rendezvous start event
         rendezvous_start_event_id = record_profiling_event(
             ProfilingEvent.RENDEZVOUS_STARTED,
-            node_id=self._this_node.addr,
+            node_id=self._this_node,
         )
 
         try:
@@ -1372,7 +1372,7 @@ def next_rendezvous(self) -> Union[RendezvousInfo, Tuple[Store, int, int]]:
         # Record rendezvous completion event
         rendezvous_completion_event_id = record_profiling_event(
             ProfilingEvent.RENDEZVOUS_COMPLETED,
-            node_id=self._this_node.addr,
+            node_id=self._this_node,
         )
 
         # Use RendezvousInfo if available (newer PyTorch versions >= 2.4.0)

src/nvidia_resiliency_ext/fault_tolerance/launcher.py

@@ -327,7 +327,7 @@ def _invoke_run_with_any_failed_policy(self, role: str = DEFAULT_ROLE) -> RunRes
                 # Record failure detection event
                 record_profiling_event(
                     ProfilingEvent.FAILURE_DETECTED,
-                    node_id=self._node_id,
+                    node_id=self._rdzv_handler._this_node,
                     rank=self._worker_group.group_rank,
                 )
 
@@ -359,7 +359,7 @@ def _invoke_run_with_any_failed_policy(self, role: str = DEFAULT_ROLE) -> RunRes
                     # Record failure detection event
                     record_profiling_event(
                         ProfilingEvent.FAILURE_DETECTED,
-                        node_id=self._node_id,
+                        node_id=self._rdzv_handler._this_node,
                         rank=self._worker_group.group_rank,
                     )
 
@@ -606,7 +606,7 @@ async def send_close_msg():
         # Record worker termination event after shutdown is complete
         record_profiling_event(
             ProfilingEvent.WORKER_TERMINATED,
-            node_id=self._node_id,
+            node_id=self._rdzv_handler._this_node,
             rank=worker_group.group_rank,
         )
 
@@ -622,7 +622,7 @@ def _start_workers(self, worker_group: WorkerGroup) -> Dict[int, Any]:
         # Record worker start start event
         record_profiling_event(
             ProfilingEvent.WORKER_START_STARTED,
-            node_id=self._node_id,
+            node_id=self._rdzv_handler._this_node,
             rank=worker_group.group_rank,
         )
 
@@ -700,7 +700,7 @@ def _start_workers(self, worker_group: WorkerGroup) -> Dict[int, Any]:
         # Record worker start completion event
         record_profiling_event(
             ProfilingEvent.WORKER_START_COMPLETED,
-            node_id=self._node_id,
+            node_id=self._rdzv_handler._this_node,
             rank=worker_group.group_rank,
         )