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

model_id = "deepseek-ai/deepseek-coder-1.3b-base"
lora_id = "Seunggg/lora-plant"

# 加载 tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)

# 加载基础模型
base = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    offload_folder="offload/",
    torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
    trust_remote_code=True
)

# 加载 LoRA adapter
model = PeftModel.from_pretrained(
    base,
    lora_id,
    offload_folder="offload/",
    torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32
)

model.eval()

# 创建 pipeline
from transformers import pipeline
pipe = pipeline(
    "text-generation",
    model=model,
    tokenizer=tokenizer,
    device_map="auto",
    max_new_tokens=256
)

def get_sensor_data():
    try:
        sensor_response = requests.get("https://arduino-realtime.onrender.com/api/data", timeout=5)
        sensor_data = sensor_response.json().get("sensorData", None)
        return json.dumps(sensor_data, ensure_ascii=False, indent=2) if sensor_data else "暂无传感器数据"
    except Exception as e:
        return "⚠️ 获取失败：" + str(e)

def respond(user_input):
    sensor_display = get_sensor_data()
    if not user_input.strip():
        return sensor_display, "请输入植物相关的问题 😊"

    prompt = f"用户提问：{user_input}\n"
    try:
        sensor_response = requests.get("https://arduino-realtime.onrender.com/api/data", timeout=5)
        sensor_data = sensor_response.json().get("sensorData", None)
        if sensor_data:
            prompt += f"当前传感器数据：{json.dumps(sensor_data, ensure_ascii=False)}\n"
        prompt += "请用更人性化的语言生成建议，并推荐相关植物文献或资料。\n回答："

        result = pipe(prompt)
        full_output = result[0]["generated_text"]
        answer = full_output.replace(prompt, "").strip()
    except Exception as e:
        answer = f"生成建议时出错：{str(e)}"

    return sensor_display, answer

# Gradio 界面
with gr.Blocks() as demo:
    with gr.Row():
        sensor_box = gr.Textbox(label="🧪 当前传感器数据", lines=6, interactive=False)
        question_box = gr.Textbox(label="🌿 植物问题", lines=4)
    answer_box = gr.Textbox(label="🤖 回答建议", lines=8, interactive=False)

    # 用户提交提问后，更新传感器+回答
    question_box.submit(fn=respond, inputs=question_box, outputs=[sensor_box, answer_box])

    # 定时每 5 秒自动刷新传感器数据
    def update_sensor():
        return get_sensor_data()

    gr.Timer(interval=5, fn=update_sensor, outputs=sensor_box)

demo.launch()