diff --git a/python/sglang/srt/hardware_backend/npu/attention/ascend_backend.py b/python/sglang/srt/hardware_backend/npu/attention/ascend_backend.py
index 356281ab7..7bd540888 100644
--- a/python/sglang/srt/hardware_backend/npu/attention/ascend_backend.py
+++ b/python/sglang/srt/hardware_backend/npu/attention/ascend_backend.py
@@ -1587,7 +1587,9 @@ class AscendAttnBackend(AttentionBackend):
                             (q.shape[0], layer.tp_q_head_num * layer.v_head_dim)
                         )
                     else:
-                        attn_output = torch.empty_like(q)
+                        attn_output = torch.empty_like(
+                            q, memory_format=torch.contiguous_format
+                        )
 
                     use_gqa = layer.tp_q_head_num != layer.tp_k_head_num
 
@@ -2746,7 +2748,9 @@ class AscendAttnBackend(AttentionBackend):
                         (q.shape[0], layer.tp_q_head_num * layer.v_head_dim)
                     )
                 else:
-                    attn_output = torch.empty_like(q)
+                    attn_output = torch.empty_like(
+                        q, memory_format=torch.contiguous_format
+                    )
 
                 use_gqa = layer.tp_q_head_num != layer.tp_k_head_num