app.py

import gradio as gr
import os
import spaces
from transformers import AutoTokenizer, AutoModelForCausalLM, AutoConfig
from threading import Thread
import torch
import time

# Set environment variables
HF_TOKEN = os.environ.get("HF_TOKEN", None)

# Apollo system prompt
SYSTEM_PROMPT = "You are Apollo, a multilingual medical model. You communicate with people and assist them."

LICENSE = """
<div style="font-family: monospace; white-space: pre; margin-top: 20px; line-height: 1.2;">
@misc{wang2024apollo,
   title={Apollo: Lightweight Multilingual Medical LLMs towards Democratizing Medical AI to 6B People},
   author={Xidong Wang and Nuo Chen and Junyin Chen and Yan Hu and Yidong Wang and Xiangbo Wu and Anningzhe Gao and Xiang Wan and Haizhou Li and Benyou Wang},
   year={2024},
   eprint={2403.03640},
   archivePrefix={arXiv},
   primaryClass={cs.CL}
}
@misc{zheng2024efficientlydemocratizingmedicalllms,
      title={Efficiently Democratizing Medical LLMs for 50 Languages via a Mixture of Language Family Experts}, 
      author={Guorui Zheng and Xidong Wang and Juhao Liang and Nuo Chen and Yuping Zheng and Benyou Wang},
      year={2024},
      eprint={2410.10626},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2410.10626}, 
}
</div>
"""

# Apollo model options
APOLLO_MODELS = {
    "Apollo": [
        "FreedomIntelligence/Apollo-7B",
        "FreedomIntelligence/Apollo-6B",
        "FreedomIntelligence/Apollo-2B",
        "FreedomIntelligence/Apollo-0.5B",
        
    ],
    "Apollo2": [
        "FreedomIntelligence/Apollo2-7B",
        "FreedomIntelligence/Apollo2-3.8B",
        "FreedomIntelligence/Apollo2-2B",
    ],
    "Apollo-MoE": [
        "FreedomIntelligence/Apollo-MoE-7B",
        "FreedomIntelligence/Apollo-MoE-1.5B",
        "FreedomIntelligence/Apollo-MoE-0.5B",
        
    ]
}

# CSS styles
css = """
h1 {
  text-align: center;
  display: block;
}
.gradio-container {
  max-width: 1200px;
  margin: auto;
}
"""

# Global variables to store currently loaded model and tokenizer
current_model = None
current_tokenizer = None
current_model_path = None

@spaces.GPU(duration=120)
def load_model(model_path, progress=gr.Progress()):
    """Load the selected model and tokenizer"""
    global current_model, current_tokenizer, current_model_path
    
    # If the same model is already loaded, don't reload it
    if current_model_path == model_path and current_model is not None:
        return "Model already loaded, no need to reload."
    
    # Clean up previously loaded model (if any)
    if current_model is not None:
        del current_model
        del current_tokenizer
        torch.cuda.empty_cache()
    
    progress(0.1, desc=f"Starting to load model {model_path}...")
    
    try:
        progress(0.3, desc="Loading tokenizer...")
        config = AutoConfig.from_pretrained(model_path, trust_remote_code=True)
        if 'MoE' in model_path:
            from configuration_upcycling_qwen2_moe import UpcyclingQwen2MoeConfig
            config = UpcyclingQwen2MoeConfig.from_pretrained(model_path, trust_remote_code=True)
            # config_moe.auto_map["AutoConfig"] = "./configuration_upcycling_qwen2_moe.UpcyclingQwen2MoeConfig"
            # config_moe.auto_map["AutoModelForCausalLM"] = "./modeling_upcycling_qwen2_moe.UpcyclingQwen2MoeForCausalLM"
        current_tokenizer = AutoTokenizer.from_pretrained(model_path, use_fast=False,trust_remote_code=True)
        
        progress(0.5, desc="Loading model...")
        if 'MoE' in model_path:
            from modeling_upcycling_qwen2_moe import UpcyclingQwen2MoeForCausalLM
            current_model = UpcyclingQwen2MoeForCausalLM.from_pretrained(
                model_path, 
                device_map="auto", 
                torch_dtype=torch.float16,
                config=config,
                trust_remote_code=True
            )
        else:
            current_model = AutoModelForCausalLM.from_pretrained(
                model_path, 
                device_map="auto", 
                torch_dtype=torch.float16,
                config=config,
                trust_remote_code=True
            )
        
        current_model_path = model_path
        progress(1.0, desc="Model loading complete!")
        return f"Model {model_path} successfully loaded."
    except Exception as e:
        progress(1.0, desc="Model loading failed!")
        return f"Model loading failed: {str(e)}"

@spaces.GPU(duration=120)
def generate_response_non_streaming(instruction, model_name, temperature=0.7, max_tokens=1024):
    """Generate a response from the Apollo model (non-streaming)"""
    global current_model, current_tokenizer, current_model_path
    print("instruction:",instruction)
    # If model is not yet loaded, load it first
    if current_model_path != model_name or current_model is None:
        load_message = load_model(model_name)
        if "failed" in load_message.lower():
            return load_message

    try:
        # 直接使用简单的提示格式，不使用模型的聊天模板
        prompt = f"User:{instruction}\nAssistant:"
        print("prompt:",prompt)
        chat_input = current_tokenizer.encode(prompt, return_tensors="pt").to(current_model.device)
        
        # 生成响应
        output = current_model.generate(
            input_ids=chat_input,
            max_new_tokens=max_tokens,
            temperature=temperature,
            do_sample=(temperature > 0),
            eos_token_id=current_tokenizer.eos_token_id  # 使用<|endoftext|>作为停止标记
        )
        
        # 解码并返回生成的文本
        generated_text = current_tokenizer.decode(output[0][len(chat_input[0]):], skip_special_tokens=True)
        print("generated_text:",generated_text)
        return generated_text
    except Exception as e:
        return f"生成响应时出错: {str(e)}"

    # try:
    #     # 检查模型是否有聊天模板
    #     if hasattr(current_tokenizer, 'chat_template') and current_tokenizer.chat_template:
    #         # 使用模型的聊天模板
    #         messages = [
    #             {"role": "system", "content": SYSTEM_PROMPT},
    #             {"role": "user", "content": instruction}
    #         ]
            
    #         # 使用模型的聊天模板格式化输入
    #         chat_input = current_tokenizer.apply_chat_template(
    #             messages, 
    #             tokenize=True, 
    #             return_tensors="pt"
    #         ).to(current_model.device)
    #     else:
    #         # 使用指定的提示格式
    #         prompt = f"User:{instruction}\nAssistant:"
    #         chat_input = current_tokenizer.encode(prompt, return_tensors="pt").to(current_model.device)
            
    #         # 获取<|endoftext|>的token id，用于停止生成
    #         eos_token_id = current_tokenizer.eos_token_id
        
    #     # 生成响应
    #     output = current_model.generate(
    #         input_ids=chat_input,
    #         max_new_tokens=max_tokens,
    #         temperature=temperature,
    #         do_sample=(temperature > 0),
    #         eos_token_id=current_tokenizer.eos_token_id  # 使用<|endoftext|>作为停止标记
    #     )
        
    #     # 解码并返回生成的文本
    #     generated_text = current_tokenizer.decode(output[0][len(chat_input[0]):], skip_special_tokens=True)
    #     return generated_text
    # except Exception as e:
    #     return f"生成响应时出错: {str(e)}"

def update_chat_with_response(chatbot, instruction, model_name, temperature, max_tokens):
    """Updates the chatbot with non-streaming response"""
    global current_model, current_tokenizer, current_model_path
    
    # If model is not yet loaded, load it first
    if current_model_path != model_name or current_model is None:
        load_result = load_model(model_name)
        if "failed" in load_result.lower():
            new_chat = list(chatbot)
            new_chat[-1] = (instruction, load_result)
            return new_chat
    
    # Generate response using the non-streaming function
    response = generate_response_non_streaming(instruction, model_name, temperature, max_tokens)
    
    # Create a copy of the current chatbot and add the response
    new_chat = list(chatbot)
    new_chat[-1] = (instruction, response)
    
    return new_chat

def on_model_series_change(model_series):
    """Update available model list based on selected model series"""
    if model_series in APOLLO_MODELS:
        return gr.update(choices=APOLLO_MODELS[model_series], value=APOLLO_MODELS[model_series][0])
    return gr.update(choices=[], value=None)

def process_message(message, chat_history, model_series_value, model_name_value, temperature_value, max_tokens_value):
    """Process user message and generate response"""
    if message.strip() == "":
        return "", chat_history
    
    # 打印用户提交的消息，用于调试
    print("instruction:", message)
    
    # Add user message to chat history
    chat_history = list(chat_history)
    chat_history.append((message, None))
    
    # 自动加载模型（如果需要）
    global current_model, current_tokenizer, current_model_path
    if current_model_path != model_name_value or current_model is None:
        try:
            load_result = load_model(model_name_value)
            if "failed" in load_result.lower():
                chat_history[-1] = (message, f"模型加载失败: {load_result}")
                return "", chat_history
        except Exception as e:
            chat_history[-1] = (message, f"模型加载出错: {str(e)}")
            return "", chat_history
    
    # Generate response
    try:
        response = generate_response_non_streaming(message, model_name_value, temperature_value, max_tokens_value)
        # Add response to chat history
        chat_history[-1] = (message, response)
    except Exception as e:
        chat_history[-1] = (message, f"生成响应时出错: {str(e)}")
    
    return "", chat_history

# Create Gradio interface
with gr.Blocks(css=css) as demo:
    # Title and description
    favicon = "🩺"
    gr.Markdown(
        f"""# {favicon} Apollo Playground
        This is a demo of the multilingual medical model series **[Apollo](https://github.com/FreedomIntelligence/Apollo)** made by **[FreedomIntelligence](https://huggingface.co/FreedomIntelligence)**.
        [Apollo1](https://arxiv.org/abs/2403.03640) supports 6 languages. [Apollo2](https://arxiv.org/abs/2410.10626) and [Apollo-MoE](https://arxiv.org/abs/2410.10626) supports 50 languages.
        """
    )
    
    with gr.Row():
        with gr.Column(scale=1):
            # Model selection controls
            model_series = gr.Dropdown(
                choices=list(APOLLO_MODELS.keys()),
                value="Apollo",
                label="Select Model Series",
                info="First choose Apollo, Apollo2 or Apollo-MoE"
            )
            
            model_name = gr.Dropdown(
                choices=APOLLO_MODELS["Apollo"],
                value=APOLLO_MODELS["Apollo"][0],
                label="Select Model Size",
                info="Select the specific model size based on the chosen model series"
            )
            
            # Parameter settings
            with gr.Accordion("Generation Parameters", open=False):
                temperature = gr.Slider(
                    minimum=0.0, 
                    maximum=1.0, 
                    value=0.7, 
                    step=0.05, 
                    label="Temperature"
                )
                max_tokens = gr.Slider(
                    minimum=128, 
                    maximum=2048, 
                    value=1024, 
                    step=32, 
                    label="Maximum Tokens"
                )
            
            # 移除Load Model按钮和状态显示
            # load_button = gr.Button("Load Model")
            # model_status = gr.Textbox(label="Model Status", value="No model loaded yet")
        
        with gr.Column(scale=2):
            # Chat interface
            chatbot = gr.Chatbot(label="Conversation", height=500, value=[])  # Initialize with empty list
            user_input = gr.Textbox(
                label="Input Medical Question",
                placeholder="Example: What are the symptoms of hypertension? 高血压有哪些症状？",
                lines=3
            )
            submit_button = gr.Button("Submit")
            clear_button = gr.Button("Clear Chat")
    
    # Event handling
    # Update model selection when model series changes
    model_series.change(
        fn=on_model_series_change,
        inputs=model_series,
        outputs=model_name
    )
    
    # 修改提交事件绑定
    submit_event = user_input.submit(
        fn=process_message,
        inputs=[user_input, chatbot, model_series, model_name, temperature, max_tokens],
        outputs=[user_input, chatbot]
    )

    submit_button.click(
        fn=process_message,
        inputs=[user_input, chatbot, model_series, model_name, temperature, max_tokens],
        outputs=[user_input, chatbot]
    )
    
    # Clear chat
    clear_button.click(
        fn=lambda: [],
        outputs=chatbot
    )
    
    # # Handle message submission
    # def user_message_submitted(message, chat_history):
    #     """Handle user submitted message"""
    #     # Ensure chat_history is a list
    #     if chat_history is None:
    #         chat_history = []
            
    #     if message.strip() == "":
    #         return "", chat_history
        
    #     # Add user message to chat history
    #     chat_history = list(chat_history)
    #     chat_history.append((message, None))
    #     return "", chat_history
    
    # # Bind message submission
    # submit_event = user_input.submit(
    #     fn=user_message_submitted,
    #     inputs=[user_input, chatbot],
    #     outputs=[user_input, chatbot]
    # ).then(
    #     fn=update_chat_with_response,
    #     inputs=[chatbot, user_input, model_name, temperature, max_tokens],
    #     outputs=chatbot
    # )
    
    # submit_button.click(
    #     fn=user_message_submitted,
    #     inputs=[user_input, chatbot],
    #     outputs=[user_input, chatbot]
    # ).then(
    #     fn=update_chat_with_response,
    #     inputs=[chatbot, user_input, model_name, temperature, max_tokens],
    #     outputs=chatbot
    # )
    
    # # Clear chat
    # clear_button.click(
    #     fn=lambda: [],
    #     outputs=chatbot
    # )
    
    examples = [
        ["Últimamente tengo la tensión un poco alta, ¿cómo debo adaptar mis hábitos?"],
        ["What are the common side effects of metformin?"],
        ["中医和西医在治疗高血压方面有什么不同的观点？"],
        ["मेरा सिर दर्द कर रहा है, मुझे क्या करना चाहिए? "],
        ["Comment savoir si je suis diabétique ?"],
        ["ما الدواء الذي يمكنني تناوله إذا لم أستطع النوم ليلاً؟"],
        ["针对一名28岁女性患者，她左小腿挫伤12小时，伤口有分泌物，骨折端外露，小腿成角畸形，描述她的最佳处理方法。"]
    ]
    gr.Examples(
        examples=examples,
        inputs=user_input
    )
    gr.HTML(LICENSE)

if __name__ == "__main__":
    demo.launch()