Skip to content

Commit 708bca1

Browse files
zxfan-cpuyangw-dev
authored andcommitted
[Bugfix] fix use-ep bug to enable ep by dp/tp size > 1 (vllm-project#16161)
Signed-off-by: Yang Wang <[email protected]>
1 parent 1084d5e commit 708bca1

File tree

1 file changed

+1
-1
lines changed
  • vllm/model_executor/layers/fused_moe

1 file changed

+1
-1
lines changed

vllm/model_executor/layers/fused_moe/layer.py

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -437,7 +437,7 @@ def __init__(
437437
# Use expert parallelism instead of tensor parallelism?
438438
vllm_config = get_current_vllm_config()
439439
use_ep = (vllm_config.parallel_config.enable_expert_parallel
440-
and self.tp_size > 1)
440+
and self.tp_size * self.dp_size > 1)
441441

442442
# For smuggling this layer into the fused moe custom op
443443
self.use_direct_call = self.dp_size == 1

0 commit comments

Comments
 (0)