Update app.py

app.py

@@ -56,11 +56,11 @@ with init_empty_weights():
 # 
 
 
-# device_map = infer_auto_device_map(model, max_memory={0: "80GB", 1: "80GB", "cpu": "65GB"})
+device_map = infer_auto_device_map(model, max_memory={0: "80GB", 1: "80GB", "cpu": "65GB"})
 
 # Load the model with the inferred device map
 # model = load_checkpoint_and_dispatch(model, model_id, device_map=device_map, no_split_module_classes=["GPTJBlock"])
-model.half()
+# model.half()