xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-05-31 14:27:08 +08:00
Code Issues Packages Projects Releases Wiki Activity

update

Browse Source 
Signed-off-by: Tyler Michael Smith <tyler@neuralmagic.com>
This commit is contained in:
Tyler Michael Smith 2025-07-10 18:27:59 -04:00
parent 29596317b0
commit df866cfebf
1 changed files with 1 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
vllm/distributed/parallel_state.py
View File

@ -782,7 +782,6 @@ class GroupCoordinator:
def prepare_communication_buffer_for_model(self, model: torch.nn.Module): def prepare_communication_buffer_for_model(self, model: torch.nn.Module):
if self.device_communicator is not None: if self.device_communicator is not None:
torch.distributed.barrier(self.device_communicator)
self.device_communicator.prepare_communication_buffer_for_model( self.device_communicator.prepare_communication_buffer_for_model(
model) model)
@ -1116,6 +1115,7 @@ def prepare_communication_buffer_for_model(model: torch.nn.Module):
MoE all2all (DeepEP) usually allocate the communication buffer MoE all2all (DeepEP) usually allocate the communication buffer
based on the model shape for optimal performance. based on the model shape for optimal performance.
""" """
get_world_group().barrier()
if _TP is not None: if _TP is not None:
_TP.prepare_communication_buffer_for_model(model) _TP.prepare_communication_buffer_for_model(model)
if _PP is not None: if _PP is not None: