diff --git a/slime/ray/rollout.py b/slime/ray/rollout.py
index cea293a508..cd501bd10d 100644
--- a/slime/ray/rollout.py
+++ b/slime/ray/rollout.py
@@ -414,6 +414,7 @@ def _try_ci_fault_injection(self):
     def dispose(self):
         for monitor in self._health_monitors:
             monitor.stop()
+        logging_utils.finish_tracking(self.args)
 
     @property
     def server(self) -> RolloutServer | None:
diff --git a/slime/utils/logging_utils.py b/slime/utils/logging_utils.py
index 5fc0fad357..1fc3d94b23 100644
--- a/slime/utils/logging_utils.py
+++ b/slime/utils/logging_utils.py
@@ -31,6 +31,16 @@ def init_tracking(args, primary: bool = True, **kwargs):
         wandb_utils.init_wandb_secondary(args, **kwargs)
 
 
+def finish_tracking(args):
+    if not args.use_wandb:
+        return
+    try:
+        if wandb.run is not None:
+            wandb.finish()
+    except Exception:
+        logging.getLogger(__name__).exception("Failed to finish wandb run")
+
+
 # TODO further refactor, e.g. put TensorBoard init to the "init" part
 def log(args, metrics, step_key: str):
     if args.use_wandb:
diff --git a/train.py b/train.py
index 15b3ca8b24..554b26b453 100644
--- a/train.py
+++ b/train.py
@@ -2,7 +2,7 @@
 
 from slime.ray.placement_group import create_placement_groups, create_rollout_manager, create_training_models
 from slime.utils.arguments import parse_args
-from slime.utils.logging_utils import configure_logger, init_tracking
+from slime.utils.logging_utils import configure_logger, init_tracking, finish_tracking
 from slime.utils.misc import should_run_periodic_action
 
 
@@ -98,6 +98,7 @@ def save(rollout_id):
             ray.get(rollout_manager.eval.remote(rollout_id))
 
     ray.get(rollout_manager.dispose.remote())
+    finish_tracking(args)
 
 
 if __name__ == "__main__":
diff --git a/train_async.py b/train_async.py
index 182309cb08..17f9cff47d 100644
--- a/train_async.py
+++ b/train_async.py
@@ -2,7 +2,7 @@
 
 from slime.ray.placement_group import create_placement_groups, create_rollout_manager, create_training_models
 from slime.utils.arguments import parse_args
-from slime.utils.logging_utils import configure_logger, init_tracking
+from slime.utils.logging_utils import configure_logger, init_tracking, finish_tracking
 from slime.utils.misc import should_run_periodic_action
 
 
@@ -72,6 +72,7 @@ def train(args):
             ray.get(rollout_manager.eval.remote(rollout_id))
 
     ray.get(rollout_manager.dispose.remote())
+    finish_tracking(args)
 
 
 if __name__ == "__main__":