Update inference.py

inference.py

@@ -357,7 +357,7 @@ def load_model(ckpt_dir, device):
     with fabric.init_module(empty_init=False):
         model = GPT(config)
 
-    model = fabric.setup(model)
+    # model = fabric.setup(model)
     state_dict = lazy_load(ckpt_dir + "/lit_model.pth")
     model.load_state_dict(state_dict, strict=True)
     model.to(device).eval()