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 10, 12:45 UTC
[Misc] Support qwen3-next lora (#31719)
BingjiaWang committed
Jan 8, 2026 at 17:27 UTC
96fcd3c267a0b16e3f94abd59e05cb3708a67db8
1 file changed
+7
-1
vllm/model_executor/models/qwen3_next.py
+7
-1
@@ -145,7 +145,13 @@ class Qwen3NextSparseMoeBlock(nn.Module):
145
prefix=f"{prefix}.gate",
146
)
147
148
- self.shared_expert_gate = torch.nn.Linear(config.hidden_size, 1, bias=False)
148
+ self.shared_expert_gate = ReplicatedLinear(
149
+ config.hidden_size,
150
+ 1,
151
+ bias=False,
152
+ quant_config=None,
153
+ prefix=f"{prefix}.shared_expert_gate",
154
+ )
155
156
if config.shared_expert_intermediate_size > 0:
157
self.shared_expert = Qwen3NextMLP(