[Reland] Make checkpoint purge thread to be a daemon thread (#978)

fegin · web-flow · commit a30d4cca690d · 2025-03-17T15:51:56.000-07:00
Reland #972 due to ghstack issues.
diff --git a/torchtitan/components/checkpoint.py b/torchtitan/components/checkpoint.py
@@ -280,7 +280,7 @@ def load_state_dict(state_dict):
                 )
             self.purge_queue = queue.Queue()
             self.purge_thread = threading.Thread(
-                target=purge_thread, args=(self.purge_queue,)
+                target=purge_thread, args=(self.purge_queue,), daemon=True
             )
             self.purge_thread.start()
         else:

Original file line number	Diff line number	Diff line change
`@@ -280,7 +280,7 @@ def load_state_dict(state_dict):`
`280`	`280`	`)`
`281`	`281`	`self.purge_queue = queue.Queue()`
`282`	`282`	`self.purge_thread = threading.Thread(`
`283`		`- target=purge_thread, args=(self.purge_queue,)`
	`283`	`+ target=purge_thread, args=(self.purge_queue,), daemon=True`
`284`	`284`	`)`
`285`	`285`	`self.purge_thread.start()`
`286`	`286`	`else:`