app.py

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

# 初始化
peft_model_id = "CMLM/ZhongJing-2-1_8b"
base_model_id = "Qwen/Qwen1.5-1.8B-Chat"

device = "cuda"
model = AutoModelForCausalLM.from_pretrained(base_model_id, device_map={"": device}).to(device)
model.load_adapter(peft_model_id)
tokenizer = AutoTokenizer.from_pretrained(
    "CMLM/ZhongJing-2-1_8b",
    padding_side="right",
    trust_remote_code=True,
    pad_token=''
)

#多轮对话
@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
    formatted_history = []
    tempuser = ""
    for entry in chat_history:
        if entry['role'] == 'user':
            tempuser = entry['content']
        elif entry['role'] == 'assistant':
            formatted_history.append((tempuser, entry['content']))
    
    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=6):
            submit_button = gr.Button("发送")
    
    submit_button.click(multi_turn_chat, [user_input, state], [chatbot, state])
    user_input.submit(multi_turn_chat, [user_input, state], [chatbot, state])
    clear_button = gr.Button("清除对话历史")
    clear_button.click(clear_history, [], [chatbot, state])

multi_turn_interface.launch()