elastic_ep: Fix stateless group port races (#36330)

Signed-off-by: Itay Alroy <ialroy@nvidia.com>
This commit is contained in:
Itay Alroy
2026-03-18 14:36:18 +00:00
committed by GitHub
parent 99267c23ca
commit de1a86b7de
12 changed files with 221 additions and 222 deletions
@@ -162,10 +162,8 @@ class ElasticEPScalingExecutor:
new_dp_size=new_dp_size,
new_world_size_across_dp=new_world_size_across_dp,
master_ip=reconfig_request.new_data_parallel_master_ip,
world_group_ports=reconfig_request.new_stateless_world_group_port_list,
dp_group_ports=reconfig_request.new_stateless_dp_group_port_list,
ep_group_ports=reconfig_request.new_stateless_ep_group_port_list,
eplb_group_ports=reconfig_request.new_stateless_eplb_group_port_list,
coord_store_port=reconfig_request.coord_store_port,
enable_eplb=updated_config.parallel_config.enable_eplb,
)
self.worker.model_runner.eep_eplb_suppressed = True
standby_ep_group = get_standby_ep_group()
+1 -12
View File
@@ -563,15 +563,4 @@ class ElasticEPScalingState:
parallel_config._data_parallel_master_port_list = (
reconfig_request.new_data_parallel_master_port_list
)
parallel_config._stateless_world_group_port_list = (
reconfig_request.new_stateless_world_group_port_list
)
parallel_config._stateless_dp_group_port_list = (
reconfig_request.new_stateless_dp_group_port_list
)
parallel_config._stateless_ep_group_port_list = (
reconfig_request.new_stateless_ep_group_port_list
)
parallel_config._stateless_eplb_group_port_list = (
reconfig_request.new_stateless_eplb_group_port_list
)
parallel_config._coord_store_port = reconfig_request.coord_store_port
+15 -9
View File
@@ -38,10 +38,8 @@ def create_standby_groups(
new_dp_size: int,
new_world_size_across_dp: int,
master_ip: str,
world_group_ports: list[list[int]],
dp_group_ports: list[list[int]],
ep_group_ports: list[list[int]],
eplb_group_ports: list[list[int]] | None = None,
coord_store_port: int,
enable_eplb: bool = True,
backend: str | None = None,
) -> None:
global \
@@ -51,19 +49,23 @@ def create_standby_groups(
_STANDBY_EP, \
_STANDBY_EPLB
from vllm.distributed.utils import get_cached_tcp_store_client
assert new_world_size_across_dp == torch.distributed.get_world_size() * new_dp_size
world_group = get_world_group()
assert isinstance(world_group, StatelessGroupCoordinator)
backend = backend or world_group.backend
coord_store = get_cached_tcp_store_client(master_ip, coord_store_port)
standby_world_ranks = [list(range(new_world_size_across_dp))]
_STANDBY_WORLD = _init_stateless_group(
standby_world_ranks,
"world",
world_group_ports,
master_ip,
backend,
use_device_communicator=False,
coord_store=coord_store,
)
_STANDBY_WORLD_NODE_COUNT = _node_count(_STANDBY_WORLD.tcp_store_group)
@@ -76,7 +78,7 @@ def create_standby_groups(
standby_dp_ranks = all_ranks.transpose(1, 3).reshape(-1, new_dp_size).unbind(0)
standby_dp_ranks = [x.tolist() for x in standby_dp_ranks]
_STANDBY_DP = _init_stateless_group(
standby_dp_ranks, "dp", dp_group_ports, master_ip, backend
standby_dp_ranks, "dp", master_ip, backend, coord_store=coord_store
)
standby_ep_ranks = (
@@ -84,12 +86,16 @@ def create_standby_groups(
)
standby_ep_ranks = [x.tolist() for x in standby_ep_ranks]
_STANDBY_EP = _init_stateless_group(
standby_ep_ranks, "ep", ep_group_ports, master_ip, backend
standby_ep_ranks, "ep", master_ip, backend, coord_store=coord_store
)
if eplb_group_ports is not None:
if enable_eplb:
_STANDBY_EPLB = _init_stateless_group(
standby_ep_ranks, "eplb", eplb_group_ports, master_ip, backend
standby_ep_ranks,
"eplb",
master_ip,
backend,
coord_store=coord_store,
)