[Elastic EP] Async preparation (#47288)

Signed-off-by: Itay Alroy <ialroy@nvidia.com>
This commit is contained in:
Itay Alroy
2026-07-28 05:19:13 -07:00
committed by GitHub
parent d2bfc6fe20
commit 35efdf6b34
21 changed files with 701 additions and 677 deletions
@@ -175,6 +175,7 @@ class DeviceCommunicatorBase:
unique_name: str = "",
global_ranks: list[int] | None = None,
global_world_size: int | None = None,
use_all2all: bool = False,
):
self.device = device or torch.device("cpu")
self.cpu_group = cpu_group
@@ -204,26 +205,15 @@ class DeviceCommunicatorBase:
self.global_world_size = dist.get_world_size()
self.rank_in_group = dist.get_group_rank(self.cpu_group, self.global_rank)
use_ep = False
all2all_backend = None
from vllm.config import get_current_vllm_config_or_none
config = get_current_vllm_config_or_none()
if config is not None:
# initialize the all2all manager for DP or sequence-parallel EP.
parallel_config = config.parallel_config
use_ep = (
parallel_config.data_parallel_size > 1
or parallel_config.use_sequence_parallel_moe
or (
parallel_config.enable_expert_parallel
and parallel_config.prefill_context_parallel_size > 1
)
)
all2all_backend = parallel_config.all2all_backend
all2all_backend = config.parallel_config.all2all_backend
self.is_ep_communicator = unique_name.split(":")[0] == "ep"
self.use_all2all = self.is_ep_communicator and use_ep
self.use_all2all = self.is_ep_communicator and use_all2all
self.all2all_backend = all2all_backend
self.all2all_manager: All2AllManagerBase | None = None
@@ -24,8 +24,11 @@ class CpuCommunicator(DeviceCommunicatorBase):
device: torch.device | None = None,
device_group: ProcessGroup | None = None,
unique_name: str = "",
use_all2all: bool = False,
):
super().__init__(cpu_group, device, device_group, unique_name)
super().__init__(
cpu_group, device, device_group, unique_name, use_all2all=use_all2all
)
self.dist_module = torch.distributed
if (
@@ -36,6 +36,7 @@ class CudaCommunicator(DeviceCommunicatorBase):
global_ranks: list[int] | None = None,
global_world_size: int | None = None,
tcp_store_group: StatelessProcessGroup | None = None,
use_all2all: bool = False,
):
super().__init__(
cpu_group,
@@ -44,6 +45,7 @@ class CudaCommunicator(DeviceCommunicatorBase):
unique_name,
global_ranks,
global_world_size,
use_all2all=use_all2all,
)
if "tp" not in unique_name:
# custom allreduce or torch symm mem can be used only by tp
@@ -20,8 +20,11 @@ class XpuCommunicator(DeviceCommunicatorBase):
device: torch.device | None = None,
device_group: ProcessGroup | None = None,
unique_name: str = "",
use_all2all: bool = False,
):
super().__init__(cpu_group, device, device_group, unique_name)
super().__init__(
cpu_group, device, device_group, unique_name, use_all2all=use_all2all
)
self.ca_comm: None = None
if self.use_all2all:
if self.all2all_backend in ("naive", "allgather_reducescatter"):