from peft import PeftModel
from transformers import GenerationConfig, LlamaForCausalLM, LlamaTokenizer


tokenizer = LlamaTokenizer.from_pretrained("decapoda-research/llama-7b-hf")
model = LlamaForCausalLM.from_pretrained(
    "decapoda-research/llama-7b-hf",
    load_in_8bit=True,
    torch_dtype=torch.float16,
    device_map="auto"
)
model = PeftModel.from_pretrained(
    model,
    "DataAgent/llama-7b-alpaca-zh-120k",
    torch_dtype=torch.float16
)