app.py

import gradio as gr
from huggingface_hub import InferenceClient
import os
from threading import Event

hf_token = os.getenv("HF_TOKEN")

stop_event = Event()

def get_model_response(client, messages, max_tokens, temperature, top_p):
    try:
        response = client.chat_completion(
            messages,
            max_tokens=max_tokens,
            temperature=temperature,
            top_p=top_p,
            stream=True
        )
        full_response = ""
        for message in response:
            if stop_event.is_set():
                break
            token = message.choices[0].delta.content if hasattr(message.choices[0], 'delta') else message.choices[0].text
            if token:
                full_response += token
                yield full_response
        stop_event.clear()
    except Exception as e:
        yield f"모델 추론 실패: {str(e)}"

def respond(message, history, system_message, max_tokens, temperature, top_p, selected_model):
    stop_event.clear()
    try:
        client = InferenceClient(model=selected_model, token=hf_token)
        
        messages = [{"role": "system", "content": system_message}]
        messages.extend([{"role": "user" if i % 2 == 0 else "assistant", "content": m} for h in history for i, m in enumerate(h) if m])
        messages.append({"role": "user", "content": message})
        
        response = ""
        for partial_response in get_model_response(client, messages, max_tokens, temperature, top_p):
            response = partial_response
            yield "", history + [(message, response)]
        
    except Exception as e:
        yield "", history + [(message, f"오류 발생: {str(e)}")]

def stop_generation():
    stop_event.set()
    return "생성이 중단되었습니다."

def continue_writing(history, system_message, max_tokens, temperature, top_p, model):
    if not history:
        return history
    last_user_message = history[-1][0]
    last_assistant_message = history[-1][1]
    
    prompt = f"이전 대화를 계속 이어서 작성해주세요. 이전 응답: {last_assistant_message}"
    
    try:
        client = InferenceClient(model=model, token=hf_token)
        messages = [{"role": "system", "content": system_message}]
        messages.extend([{"role": "user" if i % 2 == 0 else "assistant", "content": m} for h in history for i, m in enumerate(h)])
        messages.append({"role": "user", "content": prompt})
        
        response = get_model_response(client, messages, max_tokens, temperature, top_p)
        continued_response = last_assistant_message + " " + next(response)
        
        new_history = history[:-1] + [(last_user_message, continued_response)]
        return new_history
    except Exception as e:
        return history + [("시스템", f"계속 작성 중 오류 발생: {str(e)}")]    

models = {
    "deepseek-ai/DeepSeek-Coder-V2-Instruct": "(한국회사)DeepSeek-Coder-V2-Instruct",
    "meta-llama/Meta-Llama-3.1-8B-Instruct": "Meta-Llama-3.1-8B-Instruct",
    "mistralai/Mixtral-8x7B-Instruct-v0.1": "Mixtral-8x7B-Instruct-v0.1",
    "CohereForAI/c4ai-command-r-plus": "Cohere Command-R Plus"
}

with gr.Blocks() as demo:
    chatbot = gr.Chatbot(height=600)
    
    with gr.Row():
        msg = gr.Textbox(
            scale=4,
            label="메시지 입력",
            lines=3,
            placeholder="메시지를 입력하세요. 엔터 키로 줄바꿈, Shift+Enter로 전송"
        )
    
    with gr.Row():
        send = gr.Button("전송", scale=1)
        continue_btn = gr.Button("계속 작성", scale=1)

    with gr.Row():
        regenerate = gr.Button("🔄 재생성")
        stop = gr.Button("🛑 생성 중단")
        clear = gr.Button("🗑️ 대화 내역 지우기")

    with gr.Accordion("추가 설정", open=True):
        system_message = gr.Textbox(
            value="너는 나의 최고의 비서이다.\n내가 요구하는것들을 최대한 자세하고 정확하게 답변하라.\n반드시 한글로 답변할것.",
            label="시스템 메시지",
            lines=10
        )
        max_tokens = gr.Slider(minimum=1, maximum=2000, value=500, step=100, label="최대 새 토큰 수")
        temperature = gr.Slider(minimum=0.1, maximum=2.0, value=0.7, step=0.05, label="온도")
        top_p = gr.Slider(minimum=0.1, maximum=1.0, value=0.90, step=0.05, label="Top-p (핵 샘플링)")
        model = gr.Radio(list(models.keys()), value=list(models.keys())[0], label="언어 모델 선택", info="사용할 언어 모델을 선택하세요")

    # JavaScript를 사용하여 키 입력 동작 커스터마이즈
    msg.js = """
    function(e) {
        if (e.key === 'Enter' && !e.shiftKey) {
            e.preventDefault();
            return;
        }
        if (e.key === 'Enter' && e.shiftKey) {
            e.preventDefault();
            document.getElementById('component-3').click();
        }
    }
    """

    # 이벤트 핸들러 설정
    msg.submit(lambda x: "", inputs=msg, outputs=msg)  # 엔터 키 입력 시 아무 동작 안 함
    send.click(respond, inputs=[msg, chatbot, system_message, max_tokens, temperature, top_p, model], outputs=[msg, chatbot])
    continue_btn.click(continue_writing, inputs=[chatbot, system_message, max_tokens, temperature, top_p, model], outputs=[chatbot])
    regenerate.click(lambda h, s, m, t, p, mod: respond(h[-1][0] if h else "", h[:-1], s, m, t, p, mod), inputs=[chatbot, system_message, max_tokens, temperature, top_p, model], outputs=[msg, chatbot])
    stop.click(stop_generation, inputs=[], outputs=[msg])
    clear.click(lambda: (None, None), outputs=[msg, chatbot])

if __name__ == "__main__":
    if not hf_token:
        print("경고: HF_TOKEN 환경 변수가 설정되지 않았습니다. 일부 모델에 접근할 수 없을 수 있습니다.")
    demo.launch(share=True)