app.py

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

MODEL_NAME = "openbmb/MiniCPM-V-4"

@gr.cache(allow_output_mutation=True)
def load_model():
    tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, trust_remote_code=True)
    model = AutoModelForCausalLM.from_pretrained(
        MODEL_NAME,
        torch_dtype=torch.float16,
        device_map="auto",
        trust_remote_code=True
    )
    pipeline = TextGenerationPipeline(
        model=model,
        tokenizer=tokenizer,
        device=model.device.index if torch.cuda.is_available() else -1
    )
    return pipeline

def generate(prompt: str, max_length: int = 100, top_k: int = 50, top_p: float = 0.95):
    pipe = load_model()
    output = pipe(
        prompt,
        max_length=max_length,
        do_sample=True,
        top_k=top_k,
        top_p=top_p,
        num_return_sequences=1
    )
    return output[0]["generated_text"]

with gr.Blocks() as demo:
    gr.Markdown("# MiniCPM-V-4 Text Generation Demo")

    with gr.Row():
        prompt_input = gr.Textbox(label="Prompt", placeholder="여기에 입력하세요...", lines=2)

    with gr.Row():
        max_len = gr.Slider(10, 512, value=100, step=10, label="Max Length")
        topk = gr.Slider(1, 100, value=50, step=1, label="Top-k")
        topp = gr.Slider(0.1, 1.0, value=0.95, step=0.05, label="Top-p")

    generate_btn = gr.Button("Generate")
    output_box = gr.Textbox(label="Generated Text", lines=5)

    generate_btn.click(
        fn=generate,
        inputs=[prompt_input, max_len, topk, topp],
        outputs=output_box
    )

demo.launch()