fix tcp LINGER (#45190)

python/paddle/distributed/launch/context/node.py

@@ -49,6 +49,8 @@ class Node(object):
         for _ in range(100):
             with closing(socket.socket(socket.AF_INET,
                                        socket.SOCK_STREAM)) as s:
+                s.setsockopt(socket.SOL_SOCKET, socket.SO_LINGER,
+                             struct.pack('ii', 1, 0))
                 s.bind(('', 0))
                 port = s.getsockname()[1]
                 if port in self._allocated_ports:

python/paddle/distributed/launch/controllers/collective.py

@@ -93,7 +93,7 @@ class CollectiveController(Controller):
         self.pod.replicas = self.pod_replicas()
 
         # rank will be reset when restart
-        self.pod.rank = self.ctx.args.rank
+        self.pod.rank = int(self.ctx.args.rank)
 
         port = self.ctx.node.get_free_port()
 

python/paddle/distributed/launch/controllers/master.py

@@ -102,7 +102,7 @@ class HTTPMaster(Master):
             print(" ".join(cmd))
             print("-" * 80)
 
-            if self.ctx.args.rank >= 0:
+            if int(self.ctx.args.rank) >= 0:
                 self.ctx.logger.warning(
                     "--rank set in the command may not compatible in auto mode")
 

python/paddle/distributed/launch/controllers/ps.py

@@ -111,7 +111,7 @@ class PSController(Controller):
 
     def _build_pod_with_master(self):
 
-        self.pod.rank = self.ctx.args.rank
+        self.pod.rank = int(self.ctx.args.rank)
 
         server_num = self.ctx.args.server_num or 1
         servers = [