From 3689a126c30a85cce3e314bb5ffcc8aaffc1155a Mon Sep 17 00:00:00 2001
From: Haohongxiang <86215757+haohongxiang@users.noreply.github.com>
Date: Fri, 17 Jun 2022 12:27:24 +0800
Subject: [PATCH] [Dygraph] Fix barrier bugs of ProcessGroup in Eager Mode
 (#43589)

* fix pg bugs

* update
---
 python/paddle/distributed/collective.py | 1 +
 python/paddle/distributed/parallel.py   | 2 ++
 2 files changed, 3 insertions(+)

diff --git a/python/paddle/distributed/collective.py b/python/paddle/distributed/collective.py
index b92b2a3c15d..65018e7e2e6 100644
--- a/python/paddle/distributed/collective.py
+++ b/python/paddle/distributed/collective.py
@@ -378,6 +378,7 @@ def new_group(ranks=None, backend=None):
         _group_map_by_name[group_name] = group
         _group_map[gid] = group
 
+        paddle.distributed.barrier(group=group)
         return group
 
     if not backend:
diff --git a/python/paddle/distributed/parallel.py b/python/paddle/distributed/parallel.py
index f0365cab8c8..13a027db37e 100644
--- a/python/paddle/distributed/parallel.py
+++ b/python/paddle/distributed/parallel.py
@@ -20,6 +20,7 @@ from multiprocessing import Manager  # noqa: F401
 import time
 import sys
 
+import paddle
 from paddle import compat as cpt
 
 # deprecated module import
@@ -253,6 +254,7 @@ def init_parallel_env():
         _set_group_map_by_name(_default_group_name, group)
         _set_group_map(0, group)
         parallel_helper._set_parallel_ctx(True)
+        paddle.distributed.barrier(group=group)
         return group
 
     node_num = set([i.split(":")[0] for i in parallel_env.trainer_endpoints])
-- 
GitLab