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, 04:15 UTC
[perf] Enable concurrent execution of "shared_experts" and "selected_experts" in qwen3-next (#27578)
Signed-off-by: zjy0516 <riverclouds.zhu@qq.com>
Jiangyun Zhu committed
Oct 29, 2025 at 16:12 UTC
8df98c2161e28387391b667201f8458c2bdf29f4
1 file changed
+12
-5
vllm/model_executor/models/qwen3_next.py
+12
-5
@@ -159,6 +159,7 @@ class Qwen3NextSparseMoeBlock(nn.Module):
159
160
self.experts = SharedFusedMoE(
161
shared_experts=self.shared_expert,
162
+ gate=self.gate,
163
num_experts=self.n_routed_experts,
164
top_k=config.num_experts_per_tok,
165
hidden_size=config.hidden_size,
@@ -181,11 +182,17 @@ class Qwen3NextSparseMoeBlock(nn.Module):
182
if self.is_sequence_parallel:
183
hidden_states = sequence_parallel_chunk(hidden_states)
184
184
- # router_logits: (num_tokens, n_experts)
185
- router_logits, _ = self.gate(hidden_states)
186
- final_hidden_states = self.experts(
187
- hidden_states=hidden_states, router_logits=router_logits
188
- )
185
+ if self.experts.is_internal_router:
186
+ # In this case, the gate/router runs inside the FusedMoE class
187
+ final_hidden_states = self.experts(
188
+ hidden_states=hidden_states, router_logits=hidden_states
189
+ )
190
+ else:
191
+ # router_logits: (num_tokens, n_experts)
192
+ router_logits, _ = self.gate(hidden_states)
193
+ final_hidden_states = self.experts(
194
+ hidden_states=hidden_states, router_logits=router_logits
195
+ )
196
197
if self.shared_expert is not None:
198
final_hidden_states = final_hidden_states[0] + final_hidden_states[1]