From fa0219fbf80c7d8a187cc4957fe849a4c93a7164 Mon Sep 17 00:00:00 2001
From: Chranos <826995883@qq.com>
Date: Tue, 10 Feb 2026 18:22:13 +0800
Subject: [PATCH] add qwen3_moe

---
 vllm-v0.6.2/vllm/model_executor/models/qwen3_moe.py | 12 +++++++-----
 1 file changed, 7 insertions(+), 5 deletions(-)

diff --git a/vllm-v0.6.2/vllm/model_executor/models/qwen3_moe.py b/vllm-v0.6.2/vllm/model_executor/models/qwen3_moe.py
index 26415a8..8e95c52 100644
--- a/vllm-v0.6.2/vllm/model_executor/models/qwen3_moe.py
+++ b/vllm-v0.6.2/vllm/model_executor/models/qwen3_moe.py
@@ -258,11 +258,13 @@ class Qwen3MoeAttention(nn.Module):
         k_by_head = self.k_norm(k_by_head)
         k = k_by_head.reshape(k_shape)
 
-        # MLU's forward_mlu signature is (positions, x, offsets=None),
-        # so we must call separately for q and k to avoid k being
-        # treated as offsets.
-        q = self.rotary_emb(positions, q)
-        k = self.rotary_emb(positions, k)
+        # MLU rotary_emb expects a single concatenated 3D tensor, not
+        # separate q and k (forward_mlu signature differs from forward_native).
+        qk = torch.cat([q, k], dim=-1)
+        self.rotary_emb(positions,
+                        qk.view(-1, self.num_heads + self.num_kv_heads,
+                                self.head_dim))
+        q, k = qk.split([self.q_size, self.kv_size], dim=-1)
         attn_output = self.attn(q, k, v, kv_cache, attn_metadata)
         output, _ = self.o_proj(attn_output)
         return output