app.py

import gradio as gr
from transformers import LlamaTokenizer, LlamaForCausalLM
import torch
import os  # os 모듈 임포트

model_id = 'Bllossom/llama-3.2-Korean-Bllossom-3B'

# 환경 변수에서 액세스 토큰 가져오기
hf_access_token = os.getenv('HUGGINGFACEHUB_API_TOKEN')

# 토크나이저와 모델 로드
tokenizer = LlamaTokenizer.from_pretrained(
    model_id,
    use_auth_token=hf_access_token
)
model = LlamaForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    use_auth_token=hf_access_token
)

def respond(
    message,
    history,
    system_message,
    max_tokens,
    temperature,
    top_p,
):
    # 프롬프트 생성
    prompt = system_message + "\n"
    for user_msg, bot_msg in history:
        prompt += f"User: {user_msg}\nAssistant: {bot_msg}\n"
    prompt += f"User: {message}\nAssistant:"

    # 입력 토큰화
    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

    # 모델 응답 생성
    outputs = model.generate(
        **inputs,
        max_new_tokens=max_tokens,
        temperature=temperature,
        top_p=top_p,
        do_sample=True,
        eos_token_id=tokenizer.eos_token_id,
        pad_token_id=tokenizer.eos_token_id,
    )

    # 응답 디코딩
    response = tokenizer.decode(outputs[0], skip_special_tokens=True)
    response = response[len(prompt):].strip()

    # 히스토리에 추가
    history.append((message, response))

    return history

# Gradio 인터페이스 생성
demo = gr.ChatInterface(
    fn=respond,
    additional_inputs=[
        gr.Textbox(
            value="You are a friendly Chatbot.",
            label="System message"
        ),
        gr.Slider(minimum=1, maximum=2048, value=512, step=1, label="Max new tokens"),
        gr.Slider(minimum=0.1, maximum=4.0, value=0.7, step=0.1, label="Temperature"),
        gr.Slider(
            minimum=0.1,
            maximum=1.0,
            value=0.95,
            step=0.05,
            label="Top-p (nucleus sampling)",
        ),
    ],
)

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