Upload app.py

app.py

@@ -12,12 +12,12 @@ chip_map= {
     'gpt_neox.final_layer_norm': 0,
     'embed_out': 0
  }
-model = AutoModelForCausalLM.from_pretrained(name, device_map=chip_map, torch_dtype=torch.float16, load_in_8bit=True)
+model = AutoModelForCausalLM.from_pretrained(model_name, device_map=chip_map, torch_dtype=torch.float16, load_in_8bit=True)
 #model = AutoModelForCausalLM.from_pretrained(model_name)
 
 
 def predict(input, history=[], MAX_NEW_TOKENS = 500):
-    text = "User: " + input + "\n\nChip: "
+    text = "User: " + input + "\n\nAI: "
     new_user_input_ids = tokenizer(text, return_tensors="pt").input_ids
     # bot_input_ids = torch.cat([torch.LongTensor(history), new_user_input_ids], dim=-1).to("cuda")
     bot_input_ids = torch.cat([torch.LongTensor(history), new_user_input_ids], dim=-1)