diff --git a/src/ninetoothed/jit.py b/src/ninetoothed/jit.py
index 8e2113e..4c470b9 100644
--- a/src/ninetoothed/jit.py
+++ b/src/ninetoothed/jit.py
@@ -1,5 +1,7 @@
 import importlib
 import sys
+import torch
+import os 
 
 from ninetoothed.generation import CodeGenerator
 from ninetoothed.utils import calculate_default_configs
@@ -97,13 +99,10 @@ def __call__(self):
             self._max_num_configs,
             self._prettify,
         )
-        module = import_from_path(source_file, source_file)
-        module_vars = vars(module)
-
+        
         handle = _Handle(
-            module_vars[self._kernel_name],
-            module_vars[code_generator.launch_func_name],
             source_file,
+            code_generator,
         )
 
         return handle
@@ -118,11 +117,57 @@ def import_from_path(module_name, file_path):
     return module
 
 
+def get_target_device(*args, **kwargs):
+    target_device = None
+    for arg in args:
+        if isinstance(arg, torch.Tensor):
+            target_device = arg.device
+            break
+    
+    if target_device is None:
+        for val in kwargs.values():
+            if isinstance(val, torch.Tensor):
+                target_device = val.device
+                break
+                
+    return target_device
+
+
+def convert_to_cpu(source_file_path):
+    if not os.path.exists(source_file_path):
+        raise FileNotFoundError(f"源文件不存在: {source_file_path}")
+
+    dir_name = os.path.dirname(source_file_path)
+    base_name = os.path.basename(source_file_path)
+    name, ext = os.path.splitext(base_name)
+    
+    new_file_name = f"{name}_cpu{ext}"
+    new_file_path = os.path.join(dir_name, new_file_name)
+
+    with open(source_file_path, 'r', encoding='utf-8') as f:
+        content = f.read()
+
+    new_content = content.replace("triton", "triton_cpu")
+    with open(new_file_path, 'w', encoding='utf-8') as f:
+        f.write(new_content)
+
+    return new_file_path
+
+
 class _Handle:
-    def __init__(self, kernel, launch, source):
-        self._kernel = kernel
-        self._launch = launch
+    def __init__(self, source, code_generator):
         self._source = source
+        self._code_generator = code_generator
 
     def __call__(self, *args, **kwargs):
-        return self._launch(*args, **kwargs)
+        target_device = get_target_device(*args, **kwargs)
+
+        if target_device is not None and str(target_device) == "cpu":
+            cpu_path=convert_to_cpu(self._source)
+            self._source = cpu_path
+
+        module = import_from_path(self._source, self._source)
+        module_vars = vars(module)
+        self._launch = module_vars[self._code_generator.launch_func_name]
+
+        return self._launch(*args, **kwargs)
\ No newline at end of file