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 14, 05:30 UTC
[Model][Bugfix] fix ernie45 moe 300B SharedFusedMoE output tuple (#27316)
Signed-off-by: wangyafeng <wangyafeng@baidu.com>
CSWYF3634076 committed
Oct 27, 2025 at 11:53 UTC
63b22e0dbb901b75619aa4bca2dfa1d7a71f439e
1 file changed
+2
vllm/model_executor/models/ernie45_moe.py
+2
@@ -215,6 +215,8 @@ class Ernie4_5_MoeMoE(nn.Module):
215
216
if self.has_shared_experts:
217
final_hidden_states = final_hidden_states[0] + final_hidden_states[1]
218
+ else:
219
+ final_hidden_states = final_hidden_states[1]
220
221
if self.tp_size > 1:
222
final_hidden_states = self.experts.maybe_all_reduce_tensor_model_parallel(