ENH: Speed up cli interaction (#2443)

xinference/__init__.py

@@ -26,13 +26,9 @@
 def _install():
     from xoscar.backends.router import Router
 
-    from .model import _install as install_model
-
     default_router = Router.get_instance_or_empty()
     Router.set_instance(default_router)
 
-    install_model()
-
 
 _install()
 del _install

xinference/deploy/supervisor.py

@@ -31,6 +31,10 @@
 
 logger = logging.getLogger(__name__)
 
+from ..model import _install as install_model
+
+install_model()
+
 
 async def _start_supervisor(address: str, logging_conf: Optional[Dict] = None):
     logging.config.dictConfig(logging_conf)  # type: ignore

xinference/model/__init__.py

@@ -29,3 +29,7 @@ def _install():
     image_install()
     rerank_install()
     video_install()
+
+
+_install()
+del _install

xinference/utils.py

@@ -13,9 +13,8 @@
 # limitations under the License.
 
 
-import torch
-
-
 def cuda_count():
+    import torch
+
     # even if install torch cpu, this interface would return 0.
     return torch.cuda.device_count()