diff --git a/vllm_ascend/models/qwen2_5_vl.py b/vllm_ascend/models/qwen2_5_vl.py
index 2d89d95..d1a94d1 100644
--- a/vllm_ascend/models/qwen2_5_vl.py
+++ b/vllm_ascend/models/qwen2_5_vl.py
@@ -107,7 +107,7 @@ class AscendQwen2_5_VisionAttention(Qwen2_5_VisionAttention):
             for x in (q, k, v)
         ]
 
-        context_layer = torch.torch.empty_like(q)
+        context_layer = torch.empty_like(q)
 
         # operator requires pta version >= 2.5.1
         torch_npu._npu_flash_attention_unpad(
diff --git a/vllm_ascend/models/qwen2_5_vl_without_padding.py b/vllm_ascend/models/qwen2_5_vl_without_padding.py
index 291b047..47ddd44 100644
--- a/vllm_ascend/models/qwen2_5_vl_without_padding.py
+++ b/vllm_ascend/models/qwen2_5_vl_without_padding.py
@@ -87,7 +87,7 @@ class AscendQwen2_5_VisionAttention_Without_Padding(Qwen2_5_VisionAttention):
             for x in (q, k, v)
         ]
 
-        context_layer = torch.torch.empty_like(q)
+        context_layer = torch.empty_like(q)
 
         # operator requires pta version >= 2.5.1.dev20250226
         torch_npu._npu_flash_attention_unpad(
diff --git a/vllm_ascend/models/qwen2_vl.py b/vllm_ascend/models/qwen2_vl.py
index 9d049e3..a677b06 100644
--- a/vllm_ascend/models/qwen2_vl.py
+++ b/vllm_ascend/models/qwen2_vl.py
@@ -95,7 +95,7 @@ class AscendQwen2VisionAttention(Qwen2VisionAttention):
             for x in (q, k, v)
         ]
 
-        context_layer = torch.torch.empty_like(q)
+        context_layer = torch.empty_like(q)
 
         # operator requires pta version >= 2.5.1
         torch_npu._npu_flash_attention_unpad(