[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
@@ -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