Merge remote-tracking branch 'origin/dev' into dev

models/loader/loader.py

@@ -130,13 +130,13 @@ class LoaderCheckPoint:
                     # 可传入device_map自定义每张卡的部署情况
                     if self.device_map is None:
                         if 'chatglm' in model_name.lower():
-                            device_map = self.chatglm_auto_configure_device_map(num_gpus)
+                            self.device_map = self.chatglm_auto_configure_device_map(num_gpus)
                         elif 'moss' in model_name.lower():
-                            device_map = self.moss_auto_configure_device_map(num_gpus, model_name)
+                            self.device_map = self.moss_auto_configure_device_map(num_gpus, model_name)
                         else:
-                            device_map = self.chatglm_auto_configure_device_map(num_gpus)
+                            self.device_map = self.chatglm_auto_configure_device_map(num_gpus)
 
-                    model = dispatch_model(model, device_map=device_map)
+                    model = dispatch_model(model, device_map=self.device_map)
             else:
                 print(
                     "Warning: torch.cuda.is_available() returned False.\nThis means that no GPU has been "