Update app.py

app.py

@@ -30,13 +30,13 @@ routes.get_types = get_types
 
 # App code
 
-model_name = "daekeun-ml/Llama-2-ko-DPO-13B"
+model_name = "petals-team/StableBeluga2"
 
 #petals-team/StableBeluga2
 #daekeun-ml/Llama-2-ko-DPO-13B
 #daekeun-ml/Llama-2-ko-instruct-13B
 #quantumaikr/llama-2-70b-fb16-korean
-tokenizer = AutoTokenizer.from_pretrained(model_name)
+tokenizer = AutoTokenizer.from_pretrained(model_name, use_fast=False)
 
 model = None
 
@@ -151,7 +151,7 @@ def check(model_name):
 def init():
     global model
     if check(model_name):
-        model = AutoDistributedModelForCausalLM.from_pretrained(model_name)
+        model = AutoDistributedModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.float16, low_cpu_mem_usage=True, device_map="auto")
 
 
 def chat(id, npc, text):