app.py

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

# ZeroGPU 环境会自动管理 GPU 分配，因此我们不设置 CUDA_VISIBLE_DEVICES
USE_CUDA = torch.cuda.is_available()
device = torch.device("cuda: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=''
)

@spaces.GPU
def single_turn_chat(question):
    try:
        prompt = f"Question: {question}"
        messages = [
            {"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."},
            {"role": "user", "content": prompt}
        ]
        input = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
        model_inputs = tokenizer([input], return_tensors="pt").to(device)
        print("Debug: Model inputs prepared successfully.")
        
        generated_ids = model.generate(model_inputs.input_ids, max_new_tokens=512)
        print("Debug: Model generation completed successfully.")
        
        generated_ids = [output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)]
        response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
        return response
    except Exception as e:
        print(f"Error during model invocation: {str(e)}")
        raise



@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})
    
    inputs = tokenizer.apply_chat_template(chat_history, tokenize=False, add_generation_prompt=True)
    model_inputs = tokenizer([inputs], return_tensors="pt").to(device)
    
    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)
    
    chat_history.append({"role": "assistant", "content": response})
    return chat_history

# 单轮界面
single_turn_interface = gr.Interface(
    fn=single_turn_chat,
    inputs=["text"],
    outputs="text",
    title="仲景GPT-V2-1.8B 单轮对话",
    description="Unlocking the Wisdom of Traditional Chinese Medicine with AI."
)

# 多轮界面配置与之前保持一致