Update app.py

app.py

@@ -10,6 +10,7 @@ import json
 import requests
 import math
 import spaces
+import torch
 from datetime import date
 from datetime import date, datetime, timedelta
 
@@ -26,8 +27,10 @@ tokenizer.padding_side = "right"
 
 streamer = TextStreamer(tokenizer)
 
-model = AutoModelForCausalLM.from_pretrained(model, trust_remote_code=True, token = access_token, device_map="cuda", load_in_8bit=True, offload_folder="offload/")
-model = PeftModel.from_pretrained(model, peft_model, offload_folder="offload/")
+#model = AutoModelForCausalLM.from_pretrained(model, trust_remote_code=True, token = access_token, device_map="cuda", load_in_8bit=True, offload_folder="offload/")
+#model = PeftModel.from_pretrained(model, peft_model, offload_folder="offload/")
+model = AutoModelForCausalLM.from_pretrained(model, trust_remote_code=True, token = access_token, device_map="cuda", load_in_8bit=True)
+model = PeftModel.from_pretrained(model, peft_model)
 
 model = model.eval()