diff --git a/hf_mini/utils.py b/hf_mini/utils.py
index 9fee8ac..cca09e9 100644
--- a/hf_mini/utils.py
+++ b/hf_mini/utils.py
@@ -1091,11 +1091,16 @@
 
 import re
 import time
+import torch
+from typing import Union, Dict
+from transformers import AutoTokenizer
+
 from hf_mini.filter import SensitiveInforRM
-is_security = SensitiveInforRM()
 
-def input_wrapper(code_string, later_code: str = "", path: str = "") -> str:
 
+is_security = SensitiveInforRM()
+
+def input_wrapper(tokenizer: AutoTokenizer, code_string: str, later_code: str = "", path: str = "", pad_token: str = "☺" ) -> Union[Dict,None]:
     start = time.time()
     _sequerity = True
     for i in [code_string, later_code, path]:
@@ -1104,7 +1109,7 @@ def input_wrapper(code_string, later_code: str = "", path: str = "") -> str:
             break
     print(f"Done inputs checking with {(time.time()-start) * 1000:.2f}ms", flush=True)
     if not _sequerity:
-        return ""
+        return None
 
     extension_pattern = re.compile(r"(\.\w+)$")
     p = ""
@@ -1119,4 +1124,18 @@ def input_wrapper(code_string, later_code: str = "", path: str = "") -> str:
         des = LANGUAGE_WRAPPER.get(lang, "")
         if len(des) > 0 and "<AIX-SPE>" in des:
             p = des.replace("<AIX-SPE>", f"the file path is: {path}") + "\n"
-    return f"<s>▁<AIX-SPAN-PRE>▁<AIX-SPAN-POST>{later_code}▁<AIX-SPAN-MIDDLE>{p}{code_string}"
\ No newline at end of file
+
+    # SPM
+    pad_ids =  tokenizer(pad_token, return_tensors="pt", return_token_type_ids=False)
+    pad_len = len(pad_ids["input_ids"][0])
+    pre_code_ids = tokenizer("<s>▁<AIX-SPAN-PRE>▁<AIX-SPAN-POST>", return_tensors="pt", return_token_type_ids=False)
+
+    later_code_ids = tokenizer(pad_token + later_code, return_tensors="pt", return_token_type_ids=False)
+    later_code_ids["input_ids"] = later_code_ids["input_ids"][:,pad_len:]
+    later_code_ids["attention_mask"] = later_code_ids["attention_mask"][:,pad_len:]
+
+    code_string_ids = tokenizer(f"▁<AIX-SPAN-MIDDLE>{p}{code_string}", return_tensors="pt", return_token_type_ids=False)
+    code_string_ids["input_ids"] = torch.cat([pre_code_ids["input_ids"], later_code_ids["input_ids"], code_string_ids["input_ids"]], dim = 1)
+    code_string_ids["attention_mask"] = torch.cat([pre_code_ids["attention_mask"], later_code_ids["attention_mask"], code_string_ids["attention_mask"]], dim = 1)
+
+    return code_string_ids
\ No newline at end of file
diff --git a/sess_huggingface.py b/sess_huggingface.py
index 3fb91fa..d5984a8 100755
--- a/sess_huggingface.py
+++ b/sess_huggingface.py
@@ -9,19 +9,18 @@
 model = AutoModelForCausalLM.from_pretrained("aiXcoder/aixcoder-7b-base", torch_dtype=torch.bfloat16)
 
 
-text = input_wrapper(
+inputs = input_wrapper(
+    tokenizer=tokenizer,
     code_string="# 快速排序算法",
     later_code="\n",
-    path="test.py"
+    path="test.py",
 )
 
-if len(text) == 0:
+if inputs is None:
     sys.exit()
 
-inputs = tokenizer(text, return_tensors="pt", return_token_type_ids=False)
-
 inputs = inputs.to(device)
 model.to(device)
 
 outputs = model.generate(**inputs, max_new_tokens=256)
-print(tokenizer.decode(outputs[0], skip_special_tokens=False))
+print(tokenizer.decode(outputs[0], skip_special_tokens=False))
\ No newline at end of file