A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 13, 04:30 UTC
[Model] Fix bailing_moe accuracy problem (#28277)
Signed-off-by: zhaozx-cn <zhaozx2116@163.com>
zhaozx-cn committed
Nov 14, 2025 at 21:33 UTC
433c0f86751f20dbdfdeb1a711def99b7ae3df92
1 file changed
+3
-2
vllm/model_executor/models/bailing_moe.py
+3
-2
@@ -39,7 +39,6 @@ from vllm.distributed import (
39
get_pp_group,
40
get_tensor_model_parallel_rank,
41
get_tensor_model_parallel_world_size,
42
- tensor_model_parallel_all_reduce,
42
)
43
from vllm.model_executor.layers.activation import SiluAndMul
44
from vllm.model_executor.layers.fused_moe import SharedFusedMoE
@@ -330,7 +329,9 @@ class BailingMoE(nn.Module):
329
final_hidden_states = final_hidden_states + shared_output
330
331
if self.tp_size > 1:
333
- final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
332
+ final_hidden_states = self.experts.maybe_all_reduce_tensor_model_parallel(
333
+ final_hidden_states
334
+ )
335
return final_hidden_states.view(num_tokens, hidden_size)
336
337