[Bugfix] Fix NixlEPAll2AllManager's dependency on --enable-elastic-ep to function (#44422)

Signed-off-by: fangyuchu <fangyuchu@qq.com>
Co-authored-by: Tyler Michael Smith <tyler@neuralmagic.com>

vllm/distributed/device_communicators/all2all.py

@@ -9,6 +9,7 @@ import torch.distributed as dist
 
 import vllm.envs as envs
 from vllm.distributed import get_dp_group, get_ep_group
+from vllm.distributed.utils import StatelessProcessGroup
 from vllm.forward_context import get_forward_context
 from vllm.logger import init_logger
 from vllm.utils.flashinfer import (
@@ -342,7 +343,12 @@ class NixlEPAll2AllManager(All2AllManagerBase):
     _lock = threading.RLock()
 
     def __init__(self, cpu_group, tcp_store_group=None):
-        assert tcp_store_group is not None
+        if tcp_store_group is None:
+            tcp_store_group = StatelessProcessGroup(
+                rank=cpu_group.rank(),
+                world_size=cpu_group.size(),
+                store=dist.PrefixStore("nixl_ep", cpu_group.get_group_store()),
+            )
         super().__init__(cpu_group, tcp_store_group)
 
         self.max_num_ep_ranks = envs.VLLM_NIXL_EP_MAX_NUM_RANKS