app.py

import gradio as gr
from transformers import AutoTokenizer, AutoModelForVision2Seq
from PIL import Image
import torch

# Load model and tokenizer
model_id = "Qwen/Qwen-VL"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForVision2Seq.from_pretrained(model_id, trust_remote_code=True, torch_dtype=torch.float16).to("cuda").eval()

# Inference function
def ask_qwen(image, prompt):
    query = tokenizer.from_list_format([
        {"image": image},
        {"text": prompt}
    ])
    inputs = tokenizer(query, return_tensors="pt").to("cuda", torch.float16)
    with torch.no_grad():
        output = model.generate(**inputs, max_new_tokens=128)
    answer = tokenizer.decode(output[0], skip_special_tokens=True)
    return answer.strip()

# Gradio interface
demo = gr.Interface(
    fn=ask_qwen,
    inputs=[gr.Image(type="pil"), gr.Textbox(label="Prompt")],
    outputs=gr.Textbox(label="Answer"),
    title="Qwen-VL 2.5 - Vision Language Chatbot",
    description="Chat with an image using Qwen-VL 2.5 (3B)"
)

demo.launch()