add qk contiguous

vllm/model_executor/models/qwen3.py

@@ -141,14 +141,14 @@ class Qwen3Attention(nn.Module):
         if envs.VLLM_USE_APEX_RN:
             q_by_head = self.q_norm.forward_apex(q_by_head)
         else:
-            q_by_head = self.q_norm(q_by_head)
+            q_by_head = self.q_norm(q_by_head.contiguous())
         q = q_by_head.view(q.shape)
         k_by_head = k.view(*k.shape[:-1], k.shape[-1] // self.head_dim,
                            self.head_dim)
         if envs.VLLM_USE_APEX_RN:
             k_by_head = self.k_norm.forward_apex(k_by_head)
         else:
-            k_by_head = self.k_norm(k_by_head)
+            k_by_head = self.k_norm(k_by_head.contiguous())
         k = k_by_head.view(k.shape)
         q, k = self.rotary_emb(positions, q, k)
         attn_output = self.attn(q, k, v)

vllm/model_executor/models/qwen3_moe.py

@@ -234,7 +234,7 @@ class Qwen3MoeAttention(nn.Module):
         if envs.VLLM_USE_APEX_RN:
             q_by_head = self.q_norm.forward_apex(q_by_head)
         else:
-            q_by_head = self.q_norm(q_by_head)
+            q_by_head = self.q_norm(q_by_head.contiguous())
         q = q_by_head.view(q.shape)
 
         k_by_head = k.view(*k.shape[:-1], k.shape[-1] // self.head_dim,
@@ -242,7 +242,7 @@ class Qwen3MoeAttention(nn.Module):
         if envs.VLLM_USE_APEX_RN:
             k_by_head = self.k_norm.forward_apex(k_by_head)
         else:
-            k_by_head = self.k_norm(k_by_head)
+            k_by_head = self.k_norm(k_by_head.contiguous())
         k = k_by_head.view(k.shape)
         q, k = self.rotary_emb(positions, q, k)
         attn_output = self.attn(q, k, v)