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."

models = {
    "deepseek-ai/DeepSeek-Coder-V2-Instruct": "DeepSeek-Coder-V2-Instruct",
    "CohereForAI/c4ai-command-r-plus": "Cohere Command-R Plus",
    "meta-llama/Meta-Llama-3.1-8B-Instruct": "Meta-Llama-3.1-8B-Instruct"
}

with gr.Blocks() as demo:
    chatbot = gr.Chatbot()
    
    with gr.Row():
        msg = gr.Textbox(scale=4, label="메시지 입력")
        send = gr.Button("전송", 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")

    send.click(respond, inputs=[msg, chatbot, system_message, max_tokens, temperature, top_p, model], outputs=[msg, chatbot])
    msg.submit(respond, inputs=[msg, chatbot, system_message, max_tokens, temperature, top_p, model], outputs=[msg, 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)