import gradio as gr
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel



base_model_name = "hfl/llama-3-chinese-8b-instruct-v3"
adapter_model_name = "tiankuo1111/LLAMA3-TEST"

# 加载 tokenizer
tokenizer = AutoTokenizer.from_pretrained(base_model_name)

# 加载基础模型到 CPU
base_model = AutoModelForCausalLM.from_pretrained(base_model_name, torch_dtype=torch.float32, device_map=None)

# 加载 LoRA 适配器
model = PeftModel.from_pretrained(base_model, adapter_model_name)

# 运行测试
def chat_with_model(prompt):
    inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
    with torch.no_grad():
        output = model.generate(**inputs, max_new_tokens=100)
    return tokenizer.decode(output[0], skip_special_tokens=True)

iface = gr.Interface(fn=chat_with_model, inputs="text", outputs="text", title="LoRA Model Chatbot")
iface.launch()