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 4, 18:45 UTC
[BugFix]fix Qwen3 MoE call gate twice (#40664)
Signed-off-by: Kunshang Ji <kunshang.ji@intel.com>
Kunshang Ji committed
Apr 23, 2026 at 13:04 UTC
342c58bc548f6dd38c1039fdc1c5af014ee9a268
1 file changed
+13
-5
vllm/model_executor/models/qwen3_moe.py
+13
-5
@@ -231,11 +231,19 @@ class Qwen3MoeSparseMoeBlock(nn.Module):
231
if self.is_sequence_parallel:
232
hidden_states = sequence_parallel_chunk(hidden_states)
233
234
- # router_logits: (num_tokens, n_experts)
235
- router_logits, _ = self.gate(hidden_states)
236
- final_hidden_states = self.experts(
237
- hidden_states=hidden_states, router_logits=router_logits
238
- )
234
+ if self.experts.is_internal_router:
235
+ # In this case, the gate/router runs inside the FusedMoE class
236
+ final_hidden_states = self.experts(
237
+ hidden_states=hidden_states, router_logits=hidden_states
238
+ )
239
+ else:
240
+ # Actually this will be dead code, since we always pass gate into
241
+ # FusedMoE in the current implementation. But we keep this code
242
+ # here for clarity and future flexibility.
243
+ router_logits, _ = self.gate(hidden_states)
244
+ final_hidden_states = self.experts(
245
+ hidden_states=hidden_states, router_logits=router_logits
246
+ )
247
248
if self.is_sequence_parallel:
249
final_hidden_states = tensor_model_parallel_all_gather(