diff --git a/vllm/model_executor/layers/mamba/gdn/qwen_gdn_linear_attn.py b/vllm/model_executor/layers/mamba/gdn/qwen_gdn_linear_attn.py
index bbbb1c8bd5..a57c984db4 100644
--- a/vllm/model_executor/layers/mamba/gdn/qwen_gdn_linear_attn.py
+++ b/vllm/model_executor/layers/mamba/gdn/qwen_gdn_linear_attn.py
@@ -1575,14 +1575,10 @@ class QwenGatedDeltaNetAttention(GatedDeltaNetAttention):
 
         # 3. Merge core attention output
         if spec_sequence_masks is not None and core_attn_out_non_spec is not None:
-            merged_out = torch.empty(
-                (1, num_actual_tokens, *core_attn_out_spec.shape[2:]),
-                dtype=core_attn_out_non_spec.dtype,
-                device=core_attn_out_non_spec.device,
+            core_attn_out.index_copy_(0, spec_token_indx, core_attn_out_spec.squeeze(0))
+            core_attn_out.index_copy_(
+                0, non_spec_token_indx, core_attn_out_non_spec.squeeze(0)
             )
-            merged_out.index_copy_(1, spec_token_indx, core_attn_out_spec)
-            merged_out.index_copy_(1, non_spec_token_indx, core_attn_out_non_spec)
-            core_attn_out[:num_actual_tokens] = merged_out.squeeze(0)
         elif spec_sequence_masks is not None:
             core_attn_out[:num_actual_tokens] = core_attn_out_spec.squeeze(0)
         else: