From 934ed88691b49d7a5f3bed452c09cdec00274712 Mon Sep 17 00:00:00 2001
From: Chranos <826995883@qq.com>
Date: Tue, 10 Feb 2026 18:30:48 +0800
Subject: [PATCH] add qwen3_moe

---
 .../model_executor/layers/fused_moe/layer.py     | 16 +++++++++-------
 1 file changed, 9 insertions(+), 7 deletions(-)

diff --git a/vllm-v0.6.2/vllm/model_executor/layers/fused_moe/layer.py b/vllm-v0.6.2/vllm/model_executor/layers/fused_moe/layer.py
index b4389d3..a540770 100644
--- a/vllm-v0.6.2/vllm/model_executor/layers/fused_moe/layer.py
+++ b/vllm-v0.6.2/vllm/model_executor/layers/fused_moe/layer.py
@@ -153,23 +153,25 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
 
     def forward_mlu(
         self,
+        layer: torch.nn.Module,
         x: torch.Tensor,
-        w1: torch.Tensor,
-        w2: torch.Tensor,
         router_logits: torch.Tensor,
         top_k: int,
         renormalize: bool,
         use_grouped_topk: bool,
-        num_expert_group: Optional[int],
-        topk_group: Optional[int],
+        topk_group: Optional[int] = None,
+        num_expert_group: Optional[int] = None,
+        custom_routing_function: Optional[Callable] = None,
     ) -> torch.Tensor:
         from vllm._mlu_ops import fused_moe
 
-        assert use_grouped_topk is False and num_expert_group is None and topk_group is None, \
-            f"Following params: use_grouped_topk, num_expert_group, topk_group are not support yet."
+        assert use_grouped_topk is False and num_expert_group is None \
+            and topk_group is None, \
+            "Following params: use_grouped_topk, num_expert_group, " \
+            "topk_group are not supported yet."
         return fused_moe(x,
                         router_logits,
-                        w1, w2,
+                        layer.w13_weight, layer.w2_weight,
                         None, None, # bias1, bias2
                         None, # residual
                         None, # input_smooth