From df866cfebf8beb5c9fe33d3409f4efc72d8881c6 Mon Sep 17 00:00:00 2001 From: Tyler Michael Smith Date: Thu, 10 Jul 2025 18:27:59 -0400 Subject: [PATCH] update Signed-off-by: Tyler Michael Smith --- vllm/distributed/parallel_state.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vllm/distributed/parallel_state.py b/vllm/distributed/parallel_state.py index d7bb4f3aeef15..996b8147f2041 100644 --- a/vllm/distributed/parallel_state.py +++ b/vllm/distributed/parallel_state.py @@ -782,7 +782,6 @@ class GroupCoordinator: def prepare_communication_buffer_for_model(self, model: torch.nn.Module): if self.device_communicator is not None: - torch.distributed.barrier(self.device_communicator) self.device_communicator.prepare_communication_buffer_for_model( model) @@ -1116,6 +1115,7 @@ def prepare_communication_buffer_for_model(model: torch.nn.Module): MoE all2all (DeepEP) usually allocate the communication buffer based on the model shape for optimal performance. """ + get_world_group().barrier() if _TP is not None: _TP.prepare_communication_buffer_for_model(model) if _PP is not None: