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
[Bugfix] Fix Shared Expert/Zero expert code in FusedMoE.process_chunk (#25698)
Signed-off-by: Sage Moore <sage@neuralmagic.com> Co-authored-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com>
Sage Moore committed
Sep 26, 2025 at 01:25 UTC
dfb9af20142a848b114ee2bbe3e148856f59c552
1 file changed
+3
-1
vllm/model_executor/layers/fused_moe/layer.py
+3
-1
@@ -1925,7 +1925,9 @@ class FusedMoE(CustomOp):
1925
assert self.shared_experts is None or isinstance(
1926
final_hidden_states, tuple)
1927
1928
- if isinstance(final_hidden_states, tuple):
1928
+ if self.zero_expert_num is not None and self.zero_expert_num > 0:
1929
+ assert isinstance(final_hidden_states, tuple)
1930
+ assert self.shared_experts is None
1931
final_hidden_states, zero_expert_result = final_hidden_states
1932
if zero_expert_result is not None:
1933
final_hidden_states += zero_expert_result