def _endpoint_to_world_rank_id(self, endpoint):
world_endpoints = fleet.worker_endpoints()
assert endpoint in world_endpoints, "endpoint (%s) not in worker_endpoints (%s) " % (
endpoint, fleet.world_device_ids())
return world_endpoints.index(endpoint)
def test_worker_endpoints(self):
role = role_maker.PaddleCloudRoleMaker(is_collective=True)
fleet.init(role)
print(fleet.worker_endpoints(to_string=True))
print(main_programs[local_rank])
print("local rank: ", local_rank)
print("local startup program: ", startup_programs[local_rank])
startup_program = startup_programs[local_rank]
main_program = main_programs[local_rank]
loss = Loss(Block(main_program))
optimizer = ascend_optimizer.AscendOptimizer(None, fetch_list=[])
optimizer.minimize(loss,
startup_program,
auto_dp=True,
rank_table_file=os.getenv("RANK_TABLE_FILE", None))
exe = paddle.static.Executor(paddle.CPUPlace())
exe.run(startup_program)
exe.run(main_program)
worker_endpoints = fleet.worker_endpoints()
world_device_ids = fleet.world_device_ids()
local_device_ids = fleet.local_device_ids()
local_rank = int(fleet.local_rank())
print("worker_endpoints:", worker_endpoints)
print("world_device_ids:", world_device_ids)
print("local_device_ids:", local_device_ids)
print("local_rank:", local_rank)
train(worker_endpoints, world_device_ids, local_device_ids, local_rank)
def test_worker_endpoints():
"""test_worker_endpoints"""
assert fleet.worker_endpoints() == []
print("{} ... ok".format(sys._getframe().f_code.co_name))