import spaces
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
import gradio as gr
import os

os.environ['CUDA_VISIBLE_DEVICES'] = "0,1"  
USE_CUDA = torch.cuda.is_available()
device_ids_parallel = [0]
device = torch.device("cuda:{}".format(device_ids_parallel[0]) if USE_CUDA else "cpu")

# 初始化
peft_model_id = "CMLM/ZhongJing-2-1_8b"
base_model_id = "Qwen/Qwen1.5-1.8B-Chat"
model = AutoModelForCausalLM.from_pretrained(base_model_id, device_map="auto")
model.load_adapter(peft_model_id)
tokenizer = AutoTokenizer.from_pretrained(
    "CMLM/ZhongJing-2-1_8b",
    padding_side="right", 
    trust_remote_code=True,
    pad_token=''
)

#多轮  
@spaces.GPU
def multi_turn_chat(question, chat_history=None):
    if not isinstance(question, str):
        raise ValueError("The question must be a string.")
    
    if chat_history is None or chat_history == []:
        chat_history = [{"role": "system", "content": "You are a helpful TCM medical assistant named 仲景中医大语言模型, created by 医哲未来 of Fudan University."}]
    
    chat_history.append({"role": "user", "content": question})
    
    # Apply the chat template and prepare the input 
    inputs = tokenizer.apply_chat_template(chat_history, tokenize=False, add_generation_prompt=True)
    model_inputs = tokenizer([inputs], return_tensors="pt").to(device)

    try:
        # Generate the response from the model
        outputs = model.generate(model_inputs.input_ids, max_new_tokens=512)
        generated_ids = outputs[:, model_inputs.input_ids.shape[-1]:]
        response = tokenizer.decode(generated_ids[0], skip_special_tokens=True)
    except Exception as e:
        raise RuntimeError("Error in model generation: " + str(e))
    
    # Append the assistant's response to the chat history
    chat_history.append({"role": "assistant", "content": response})
    
    # Format the chat history for output
    tempass = ""
    tempuser = ""
    formatted_history = []
    for entry in chat_history:
        if entry['role'] == 'user':
            tempuser = entry['content'] 
        elif entry['role'] == 'assistant':
            tempass = entry['content']
            temp = tempuser,tempass 
            formatted_history.append(temp)
    
    return formatted_history, chat_history


def clear_history():
    return [], []

# 多轮界面  
with gr.Blocks() as multi_turn_interface:
    chatbot = gr.Chatbot(label="仲景GPT-V2-1.8B 多轮对话")
    state = gr.State([])  
    with gr.Row():
        with gr.Column(scale=6):
            user_input = gr.Textbox(label="输入",placeholder="输入你的问题")
        with gr.Column(scale=1):
            submit_btn = gr.Button("提交")
            clear_history_btn = gr.Button("清除历史对话")
    submit_btn.click(multi_turn_chat, [user_input, state], [chatbot, state], concurrency_limit=10)
    clear_history_btn.click(fn=clear_history, inputs=None, outputs=[chatbot, state], queue=False)
    user_input.submit(multi_turn_chat, [user_input, state], [chatbot, state], concurrency_limit=10)

multi_turn_interface.launch(debug=True, server_name='0.0.0.0', server_port=6006, max_threads=200)