Update app.py

app.py

@@ -65,6 +65,7 @@ tokenizer = AutoTokenizer.from_pretrained(model_id)
 hf_save_dir = "./bitnet"
 model = AutoModelForCausalLM.from_pretrained(
     model_id,
+    torch_dtype=torch.bfloat32
     device_map="auto"
 )
 device = model.device
@@ -73,8 +74,6 @@ if torch.cuda.is_available():
 else:
     print("CUDA not available; using CPU.")
 
-
-model.to(device)
 print(f"Loaded pre-trained Hugging Face model '{model_id}'.")
 
 # ---------------------------------------------------------------------------------