num_min_workers = int(params["min_workers"])
num_max_workers = int(params.get("max_workers", num_min_workers))
assert num_min_workers >= 1, "Min number of workers should be at least 1"
assert (
num_max_workers >= num_min_workers
), "Max number of workers cannot be less than min number of workers"
timeout = int(params.get("timeout", CONST_DEFAULT_OVERALL_TIMEOUT))
last_call_timeout = int(
params.get("last_call_timeout", CONST_DEFAULT_LAST_CALL_TIMEOUT)
)
kwargs = _parse_etcd_client_params(params)
# Etcd rendezvous implementation
etcd_rdzv = CustomRendezvous(
endpoints=etcd_endpoints,
prefix=etcd_prefix,
run_id=run_id,
num_min_workers=num_min_workers,
num_max_workers=num_max_workers,
timeout=timeout,
last_call_timeout=last_call_timeout,
**kwargs,
)
return CustomRendezvousHandler(rdzv_impl=etcd_rdzv)
# torchelastic.rendezvous.RendezvousHandler using etcd (API v2):
register_rendezvous_handler("custom", _custom_rendezvous_handler)
etcd_prefix = params.get("etcd_prefix", "/torchelastic/p2p")
num_min_workers = int(params["min_workers"])
num_max_workers = int(params.get("max_workers", num_min_workers))
assert num_min_workers >= 1, "Min number of workers should be at least 1"
assert (
num_max_workers >= num_min_workers
), "Max number of workers cannot be less than min number of workers"
timeout = int(params.get("timeout", CONST_DEFAULT_OVERALL_TIMEOUT))
last_call_timeout = int(
params.get("last_call_timeout", CONST_DEFAULT_LAST_CALL_TIMEOUT))
kwargs = _parse_etcd_client_params(params)
# Etcd rendezvous implementation
etcd_rdzv = EtcdRendezvous(
endpoints=etcd_endpoints,
prefix=etcd_prefix,
run_id=run_id,
num_min_workers=num_min_workers,
num_max_workers=num_max_workers,
timeout=timeout,
last_call_timeout=last_call_timeout,
kwargs=kwargs,
)
return EtcdRendezvousHandler(rdzv_impl=etcd_rdzv)
# torchelastic.rendezvous.RendezvousHandler using etcd (API v2):
register_rendezvous_handler("etcd", _etcd_rendezvous_handler)
import torchelastic.rendezvous.parameters as parameters
# pyre-fixme[21]: Could not find name `register_rendezvous_handler` in
# `torch.distributed`.
from torch.distributed import register_rendezvous_handler
class MockedRdzv(object):
pass
def get_mock_rdzv(url):
return MockedRdzv()
register_rendezvous_handler("mocked-rdzv", get_mock_rdzv)
class RendezvousHandlerFactoryTest(unittest.TestCase):
def test_construct_rdzv_url(self):
params = parameters.RendezvousParameters(
"etcd",
"localhost:8081",
"1234",
1,
4,
"timeout=60,protocol=https,key=/etc/kubernetes/certs/client.key",
)
actual_url = parameters._construct_rendezvous_url(params)
expected_url = ("etcd://localhost:8081/1234"