[A2A] Allow EPLB + SharedExpert Overlap + DeepEPv2 (#57236)

Signed-off-by: Robert Shaw <robshaw@redhat.com>
Signed-off-by: Robert Shaw <robertgshaw2@gmail.com>
Co-authored-by: Robert Shaw <robshaw@redhat.com>
This commit is contained in:
Robert Shaw
2026-09-17 04:02:39 +00:00
committed by GitHub
co-authored by Robert Shaw
parent 68f2c7e74f
commit c38f92aae7
2 changed files with 5 additions and 0 deletions
@@ -1,5 +1,6 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import os
import threading
from collections.abc import Iterator
from contextlib import contextmanager
@@ -1048,6 +1049,9 @@ class DeepEPV2All2AllManager(All2AllManagerBase):
# missing GIN support.
probe = torch.zeros(1, device="cuda")
torch.distributed.all_reduce(probe, group=group)
# DeepEPv2 respects EP_DISABLE_GIN, so skip the GIN requirement check.
if os.environ.get("EP_DISABLE_GIN", "0") != "0":
return
gin_type = query_nccl_gin_type(group)
if gin_type is None:
@@ -85,6 +85,7 @@ class SharedExperts(torch.nn.Module):
# Both these comm backends have been shown to be safe for shared expert overlap.
_EPLB_OVERLAP_SAFE_BACKENDS = (
"allgather_reducescatter",
"deepep_v2",
"flashinfer_nvlink_one_sided",
)