未验证 提交 b0bd93de 编写于 作者: L lilong12 提交者: GitHub

Disable gloo by default (#29805)

* update, test=develop
上级 b6fd2629
...@@ -220,15 +220,8 @@ class Gloo(object): ...@@ -220,15 +220,8 @@ class Gloo(object):
rank, nodes = self._get_rank_nodes(Role.WORKER) rank, nodes = self._get_rank_nodes(Role.WORKER)
gloo = init(rank, nodes, "WORKER") gloo = init(rank, nodes, "WORKER")
self._worker_comm = gloo self._worker_comm = gloo
else: # TODO (sandyhouse): initialize gloo for server and all
rank, nodes = self._get_rank_nodes(Role.SERVER)
gloo = init(rank, nodes, "SERVER")
self._server_comm = gloo
if self._need_init_all:
rank, nodes = self._get_rank_nodes(Role.ALL)
gloo = init(rank, nodes, "ALL")
self._nodes_comm = gloo
if start_http_server: if start_http_server:
http_server_d["running"] = False http_server_d["running"] = False
http_server.join() http_server.join()
......
...@@ -219,7 +219,7 @@ def launch_collective(args): ...@@ -219,7 +219,7 @@ def launch_collective(args):
global_envs = copy.copy(os.environ.copy()) global_envs = copy.copy(os.environ.copy())
gloo_rendezvous_dir = tempfile.mkdtemp() gloo_rendezvous_dir = tempfile.mkdtemp()
# add gloo env # add gloo env
global_envs["PADDLE_WITH_GLOO"] = str(os.getenv("PADDLE_WITH_GLOO", "1")) global_envs["PADDLE_WITH_GLOO"] = str(os.getenv("PADDLE_WITH_GLOO", "0"))
global_envs["PADDLE_GLOO_RENDEZVOUS"] = "3" global_envs["PADDLE_GLOO_RENDEZVOUS"] = "3"
global_envs["PADDLE_GLOO_FS_PATH"] = gloo_rendezvous_dir global_envs["PADDLE_GLOO_FS_PATH"] = gloo_rendezvous_dir
......
...@@ -954,7 +954,7 @@ class ParameterServerLauncher(object): ...@@ -954,7 +954,7 @@ class ParameterServerLauncher(object):
"TRAINING_ROLE": "PSERVER", "TRAINING_ROLE": "PSERVER",
"PADDLE_TRAINERS_NUM": str(self.worker_num), "PADDLE_TRAINERS_NUM": str(self.worker_num),
"POD_IP": cur_server.endpoint.split(":")[0], "POD_IP": cur_server.endpoint.split(":")[0],
"PADDLE_WITH_GLOO": str(os.getenv("PADDLE_WITH_GLOO", "1")), "PADDLE_WITH_GLOO": str(os.getenv("PADDLE_WITH_GLOO", "0")),
"PADDLE_GLOO_RENDEZVOUS": "3", "PADDLE_GLOO_RENDEZVOUS": "3",
"PADDLE_GLOO_FS_PATH": self.gloo_rendezvous_dir, "PADDLE_GLOO_FS_PATH": self.gloo_rendezvous_dir,
"PADDLE_GLOO_HTTP_ENDPOINT": self.http_port "PADDLE_GLOO_HTTP_ENDPOINT": self.http_port
...@@ -1018,7 +1018,7 @@ class ParameterServerLauncher(object): ...@@ -1018,7 +1018,7 @@ class ParameterServerLauncher(object):
self.heter_worker_endpoints, self.heter_worker_endpoints,
"TRAINING_ROLE": "TRAINER", "TRAINING_ROLE": "TRAINER",
"PADDLE_TRAINER_ID": str(cur_worker.rank), "PADDLE_TRAINER_ID": str(cur_worker.rank),
"PADDLE_WITH_GLOO": str(os.getenv("PADDLE_WITH_GLOO", "1")), "PADDLE_WITH_GLOO": str(os.getenv("PADDLE_WITH_GLOO", "0")),
"PADDLE_GLOO_RENDEZVOUS": "3", "PADDLE_GLOO_RENDEZVOUS": "3",
"PADDLE_GLOO_FS_PATH": self.gloo_rendezvous_dir, "PADDLE_GLOO_FS_PATH": self.gloo_rendezvous_dir,
"FLAGS_selected_gpus": "0", "FLAGS_selected_gpus": "0",
...@@ -1088,7 +1088,7 @@ class ParameterServerLauncher(object): ...@@ -1088,7 +1088,7 @@ class ParameterServerLauncher(object):
"TRAINING_ROLE": "HETER_TRAINER", "TRAINING_ROLE": "HETER_TRAINER",
"PADDLE_TRAINERS_NUM": str(self.worker_num), "PADDLE_TRAINERS_NUM": str(self.worker_num),
"POD_IP": cur_heter_worker.endpoint.split(":")[0], "POD_IP": cur_heter_worker.endpoint.split(":")[0],
"PADDLE_WITH_GLOO": str(os.getenv("PADDLE_WITH_GLOO", "1")), "PADDLE_WITH_GLOO": str(os.getenv("PADDLE_WITH_GLOO", "0")),
"PADDLE_GLOO_RENDEZVOUS": "3", "PADDLE_GLOO_RENDEZVOUS": "3",
"PADDLE_GLOO_FS_PATH": self.gloo_rendezvous_dir, "PADDLE_GLOO_FS_PATH": self.gloo_rendezvous_dir,
"FLAGS_selected_gpus": "0", "FLAGS_selected_gpus": "0",
......
...@@ -142,6 +142,8 @@ def init_parallel_env(): ...@@ -142,6 +142,8 @@ def init_parallel_env():
_check_var_exists("PADDLE_TRAINER_ENDPOINTS") _check_var_exists("PADDLE_TRAINER_ENDPOINTS")
# 3: init gloo context (step 1: httpsever start) # 3: init gloo context (step 1: httpsever start)
init_gloo = int(os.getenv("PADDLE_WITH_GLOO", "0"))
if init_gloo:
ep_rank_0 = parallel_env.trainer_endpoints[0].split(":") ep_rank_0 = parallel_env.trainer_endpoints[0].split(":")
ep_rank = parallel_env.trainer_endpoints[parallel_env.rank].split(":") ep_rank = parallel_env.trainer_endpoints[parallel_env.rank].split(":")
manager = Manager() manager = Manager()
...@@ -185,6 +187,7 @@ def init_parallel_env(): ...@@ -185,6 +187,7 @@ def init_parallel_env():
# dividing init_gloo into two part beacause nccl and gloo # dividing init_gloo into two part beacause nccl and gloo
# are separately looking for free ports which sometimes # are separately looking for free ports which sometimes
# leads to port-conflict. # leads to port-conflict.
if init_gloo:
wait_server_ready([parallel_env.trainer_endpoints[0]]) wait_server_ready([parallel_env.trainer_endpoints[0]])
gloo_strategy = core.GlooParallelStrategy() gloo_strategy = core.GlooParallelStrategy()
......
...@@ -178,6 +178,7 @@ class TestDistBase(unittest.TestCase): ...@@ -178,6 +178,7 @@ class TestDistBase(unittest.TestCase):
"LD_PRELOAD": os.getenv("LD_PRELOAD", ""), "LD_PRELOAD": os.getenv("LD_PRELOAD", ""),
"GLOG_v": "0", "GLOG_v": "0",
"NCCL_P2P_DISABLE": "1", "NCCL_P2P_DISABLE": "1",
"PADDLE_WITH_GLOO": "1",
"BACKEND": backend, "BACKEND": backend,
"PATH_ID": path_id "PATH_ID": path_id
} }
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册