Skip to content

Commit e3dad30

Browse files
committed
fix stupid bug
Signed-off-by: Bill Nell <bnell@redhat.com>
1 parent bd1d8d4 commit e3dad30

File tree

1 file changed

+4
-1
lines changed
  • vllm/model_executor/layers/fused_moe

1 file changed

+4
-1
lines changed

vllm/model_executor/layers/fused_moe/layer.py

Lines changed: 4 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -121,12 +121,15 @@ def init_prepare_finalize(self, moe: FusedMoEConfig,
121121

122122
handle = all2all_manager.get_handle(all_to_all_args)
123123

124+
assert moe.tp_size == all2all_manager.tp_group.world_size
125+
124126
prepare_finalize = PplxPrepareAndFinalize(
125127
handle,
126128
max_num_tokens=moe.max_num_tokens,
127129
world_size=all2all_manager.world_size,
128130
rank=all2all_manager.rank,
129-
dp_size=moe.dp_size,
131+
# dp_size actually means tp_size, bug in pplx kernels
132+
dp_size=moe.tp_size,
130133
)
131134
elif moe.use_deepep_ht_kernels:
132135
assert moe.dp_size == all2all_manager.dp_world_size

0 commit comments

Comments
 (0)