diff --git a/douzero/dmc/arguments.py b/douzero/dmc/arguments.py
index 7f4b9f2..06190b5 100644
--- a/douzero/dmc/arguments.py
+++ b/douzero/dmc/arguments.py
@@ -13,19 +13,17 @@ parser.add_argument('--objective', default='adp', type=str, choices=['adp', 'wp'
 # Training settings
 parser.add_argument('--onnx_sync_interval', default=120, type=int,
                     help='Time interval (in seconds) at which to sync the onnx model')
-parser.add_argument('--actor_device_cpu', action='store_true',
-                    help='Use CPU as actor device')
 parser.add_argument('--gpu_devices', default='0', type=str,
                     help='Which GPUs to be used for training')
 parser.add_argument('--infer_devices', default='0', type=str,
                     help='Which device to be used for infer')
-parser.add_argument('--num_infer', default=3, type=int,
+parser.add_argument('--num_infer', default=2, type=int,
                     help='The number of process used for infer')
 parser.add_argument('--num_actor_devices', default=1, type=int,
                     help='The number of devices used for simulation')
-parser.add_argument('--num_actors', default=4, type=int,
+parser.add_argument('--num_actors', default=3, type=int,
                     help='The number of actors for each simulation device')
-parser.add_argument('--num_actors_thread', default=6, type=int,
+parser.add_argument('--num_actors_thread', default=4, type=int,
                     help='The number of actors for each simulation device')
 parser.add_argument('--training_device', default='0', type=str,
                     help='The index of the GPU used for training models. `cpu` means using cpu')
diff --git a/douzero/dmc/dmc.py b/douzero/dmc/dmc.py
index cf8ed83..20febab 100644
--- a/douzero/dmc/dmc.py
+++ b/douzero/dmc/dmc.py
@@ -70,9 +70,9 @@ def train(flags):
     Then it will start subprocesses as actors. Then, it will call
     learning function with  multiple threads.
     """
-    if not flags.actor_device_cpu or flags.training_device != 'cpu':
+    if flags.training_device != 'cpu' or flags.infer_devices != 'cpu':
         if not torch.cuda.is_available():
-            raise AssertionError("CUDA not available. If you have GPUs, please specify the ID after `--gpu_devices`. Otherwise, please train with CPU with `python3 train.py --actor_device_cpu --training_device cpu`")
+            raise AssertionError("CUDA not available. If you have GPUs, please specify the ID after `--gpu_devices`. Otherwise, please train with CPU with `python3 train.py --infer_devices cpu --training_device cpu`")
     plogger = FileWriter(
         xpid=flags.xpid,
         xp_args=flags.__dict__,
diff --git a/douzero/dmc/models.py b/douzero/dmc/models.py
index 89a2793..3fce8ed 100644
--- a/douzero/dmc/models.py
+++ b/douzero/dmc/models.py
@@ -493,8 +493,11 @@ model_dict_new_lite['landlord_up'] = GeneralModelLite
 model_dict_new_lite['landlord_front'] = GeneralModelLite
 model_dict_new_lite['landlord_down'] = GeneralModelLite
 
-def forward_logic(self_model, position, z, x, return_value=False, flags=None):
+def forward_logic(self_model, position, z, x, device='cpu', return_value=False, flags=None):
     legal_count = len(z)
+    if not flags.enable_onnx:
+        z = torch.tensor(z, device=device)
+        x = torch.tensor(x, device=device)
     if legal_count >= 80:
         partition_count = int(legal_count / 40)
         sub_z = np.array_split(z, partition_count)
@@ -577,8 +580,8 @@ class OldModel:
     def get_onnx_params(self, position):
         self.models[position].get_onnx_params(self.device)
 
-    def forward(self, position, z, x, return_value=False, flags=None):
-        return forward_logic(self, position, z, x, return_value, flags)
+    def forward(self, position, z, x, device='cpu', return_value=False, flags=None):
+        return forward_logic(self, position, z, x, device, return_value, flags)
 
     def share_memory(self):
         if self.models['landlord'] is not None:
@@ -646,8 +649,8 @@ class Model:
     def get_onnx_params(self, position):
         self.models[position].get_onnx_params(self.device)
 
-    def forward(self, position, z, x, return_value=False, flags=None, debug=False):
-        return forward_logic(self, position, z, x, return_value, flags)
+    def forward(self, position, z, x, device='cpu', return_value=False, flags=None):
+        return forward_logic(self, position, z, x, device, return_value, flags)
 
     def share_memory(self):
         if self.models['landlord'] is not None:
diff --git a/douzero/dmc/utils.py b/douzero/dmc/utils.py
index bb8b53f..646cbc8 100644
--- a/douzero/dmc/utils.py
+++ b/douzero/dmc/utils.py
@@ -131,7 +131,7 @@ def infer_logic(i, device, infer_queues, model, flags, onnx_frame):
             try:
                 task = infer_queue['input'].get_nowait()
                 with torch.no_grad():
-                    result = model.forward(task['position'], task['z_batch'], task['x_batch'], return_value=True, flags=flags)
+                    result = model.forward(task['position'], task['z_batch'], task['x_batch'], device=device, return_value=True, flags=flags)
                 infer_queue['output'].put({
                     'values': result['values']
                 })