add api

app.py

@@ -7,13 +7,13 @@ from pydantic import BaseModel
 from transformers import AutoTokenizer, AutoModelForCausalLM, LlamaTokenizer
 
 os.environ["CUDA_DEVICE_ORDER"]="PCI_BUS_ID"
-os.environ['CUDA_VISIBLE_DEVICES'] = "0"
+os.environ['CUDA_VISIBLE_DEVICES'] = "2"
 
 MODEL_NAME = "RegularizedSelfPlay/sppo_forward1reverse5-0.1-PromptABC-Mistral-7B-Instruct-SPPO-Iter3"  # Example: "meta-llama/Llama-2-7b-chat-hf"
 HF_TOKEN = os.getenv("HF_API_TOKEN")
 
 # Load model and tokenizer
-tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-Instruct-v0.2", token=HF_TOKEN)
+tokenizer = LlamaTokenizer.from_pretrained("mistralai/Mistral-7B-Instruct-v0.2", token=HF_TOKEN)
 llm = LLM(
         model=MODEL_NAME, 
         # revision="1296dc8fd9b21e6424c9c305c06db9ae60c03ace",