app.py

import spaces  # 必须在最顶部导入
import gradio as gr
import os

# 获取 Hugging Face 访问令牌
hf_token = os.getenv("HF_API_TOKEN")

# 定义基础模型名称
base_model_name = "larry1129/meta-llama-3.1-8b-bnb-4bit"

# 定义 adapter 模型名称
adapter_model_name = "larry1129/WooWoof_AI"

# 定义全局变量用于缓存模型和分词器
model = None
tokenizer = None

# 定义提示生成函数
def generate_prompt(instruction, input_text=""):
    if input_text:
        prompt = f"""### Instruction:
{instruction}
### Input:
{input_text}
### Response:
"""
    else:
        prompt = f"""### Instruction:
{instruction}
### Response:
"""
    return prompt

# 定义生成响应的函数，并使用 @spaces.GPU 装饰
@spaces.GPU(duration=120)
def generate_response(instruction, input_text):
    global model, tokenizer

    if model is None:
        # 检查 bitsandbytes 是否已安装
        import importlib.util
        if importlib.util.find_spec("bitsandbytes") is None:
            import subprocess
            subprocess.call(["pip", "install", "--upgrade", "bitsandbytes"])

        # 在函数内部导入需要 GPU 的库
        import torch
        from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
        from peft import PeftModel

        # 创建量化配置
        bnb_config = BitsAndBytesConfig(
            load_in_4bit=True,
            bnb_4bit_use_double_quant=True,
            bnb_4bit_quant_type="nf4",
            bnb_4bit_compute_dtype=torch.float16
        )

        # 加载分词器
        tokenizer = AutoTokenizer.from_pretrained(base_model_name, use_auth_token=hf_token)

        # 加载基础模型
        base_model = AutoModelForCausalLM.from_pretrained(
            base_model_name,
            quantization_config=bnb_config,
            device_map="auto",
            use_auth_token=hf_token,
            trust_remote_code=True
        )

        # 加载 adapter 并将其应用到基础模型上
        model = PeftModel.from_pretrained(
            base_model,
            adapter_model_name,
            torch_dtype=torch.float16,
            use_auth_token=hf_token
        )

        # 设置 pad_token
        tokenizer.pad_token = tokenizer.eos_token
        model.config.pad_token_id = tokenizer.pad_token_id

        # 切换到评估模式
        model.eval()
    else:
        # 在函数内部导入需要的库
        import torch

    # 生成提示
    prompt = generate_prompt(instruction, input_text)
    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

    with torch.no_grad():
        outputs = model.generate(
            input_ids=inputs["input_ids"],
            attention_mask=inputs.get("attention_mask"),
            max_new_tokens=128,
            temperature=0.7,
            top_p=0.95,
            do_sample=True,
        )
    response = tokenizer.decode(outputs[0], skip_special_tokens=True)
    response = response.split("### Response:")[-1].strip()
    return response

# 创建 Gradio 接口
iface = gr.Interface(
    fn=generate_response,
    inputs=[
        gr.Textbox(lines=2, placeholder="请输入指令...", label="Instruction"),
        gr.Textbox(lines=2, placeholder="如果有额外输入，请在此填写...", label="Input (可选)")
    ],
    outputs="text",
    title="WooWoof AI 交互式聊天",
    description="基于 LLAMA 3.1 的大语言模型，支持指令和可选输入。",
    allow_flagging="never"
)

# 启动 Gradio 接口
iface.launch()