import gradio as gr
from transformers import AutoTokenizer, AutoModelForCausalLM, AutoProcessor
from huggingface_hub import login
import os
import torch

HF_TOKEN = os.environ.get("HF_TOKEN")
login(token=HF_TOKEN)

MODEL_ID = "Qwen/Qwen-VL-Chat-Int4"

processor = AutoProcessor.from_pretrained(MODEL_ID, trust_remote_code=True, token=HF_TOKEN)
model = AutoModelForCausalLM.from_pretrained(MODEL_ID, trust_remote_code=True, device_map="auto", token=HF_TOKEN)
model.eval()

def ask(image, prompt):
    inputs = processor(images=image, text=prompt, return_tensors="pt").to(model.device)
    outputs = model.generate(**inputs, max_new_tokens=512)
    response = processor.batch_decode(outputs, skip_special_tokens=True)[0]
    return response

demo = gr.Interface(
    fn=ask,
    inputs=[gr.Image(type="pil"), gr.Textbox(label="請輸入問題")],
    outputs="text",
    title="🧠 Qwen1.5-VL 圖文問答 Demo"
)

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