Update app.py

app.py

@@ -37,7 +37,8 @@ tokenizer = AutoTokenizer.from_pretrained(model_id,token=access_token)
 model = AutoModelForCausalLM.from_pretrained(
     model_id,
     device_map=device,
-    torch_dtype=torch.float8,
+    #torch_dtype=torch.float8,
+    load_in_8bit=True,
     token=access_token
 )
 model.eval()